A High-Resolution Aerosol Optical Thickness Inversion Method and System Based on Graph Networks
By using a graph network-based method to preprocess and fuse features of multi-source remote sensing data from both satellite and ground, the problems of spatial heterogeneity and computational complexity in high-resolution satellite data in traditional AOD inversion methods are solved, achieving efficient high-resolution AOD inversion and improving the accuracy and efficiency of the inversion results.
Patent Information
- Application Number
- CN202410198901.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-02-22
- Publication Date
- 2026-11-17
- Estimated Expiration
- 2044-02-22
Smart Images

Figure CN118149714B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of atmospheric environment remote sensing, and in particular to a method and system for inverting high-resolution aerosol optical thickness based on graph networks. Background Technology
[0002] Traditional aerosol optical thickness (AOD) retrieval methods, represented by the dark target method and the Deep Blue algorithm, rely on the simulation of atmospheric radiative transfer processes. Accurate land-atmosphere decoupling is crucial for improving AOD retrieval quality. Therefore, to accurately solve ill-conditioned problems in AOD retrieval, the estimation of surface reflectance requires stringent assumptions and necessary prior information.
[0003] To simplify the complex process of AOD physical inversion, technicians often use machine learning methods for quantitative aerosol parameter inversion. These methods typically use long-term, high-precision aerosol station measurements as the model training target, fitting the complex nonlinear relationship between AOD and apparent / surface reflectance. In these methods, the quality of the training data directly affects the model accuracy; results for areas lacking real-world ground-based observations are highly uncertain. Furthermore, these inversion methods usually perform AOD inversion pixel-by-pixel independent calculations and perform appropriate spatial aggregation operations on the inversion results. Therefore, they cannot fully exploit the rich information of neighboring pixels during the inversion process and struggle to account for the spatial similarity and heterogeneity constraints of regional aerosol distribution.
[0004] Secondly, the spatial heterogeneity of high-resolution satellite observation data is a significant issue. Traditional physical inversion methods are more sensitive to noise in the input data, which increases the uncertainty of the inversion results. Furthermore, traditional physical inversion methods lack sufficient flexibility and adaptability when processing high-resolution data, making it difficult to account for spatial details in the inversion results. Additionally, processing large-scale sample points may also present high computational complexity and time consumption. While machine learning methods guided by aerosol station observation data can overcome these problems to some extent, there is currently a lack of effective technical methods to achieve AOD inversion tasks from high-resolution satellites. Summary of the Invention
[0005] This application discloses a high-resolution aerosol optical thickness inversion method and system based on graph networks, which is used to overcome the spatial heterogeneity problem of high-resolution observation data in the inversion process, and fully explore the information of neighboring pixels in the inversion process, taking into account the spatial similarity and heterogeneity constraints of regional aerosol distribution, so as to realize the AOD inversion task of high-resolution satellites.
[0006] The first aspect of this application provides a high-resolution aerosol optical thickness inversion method based on graph networks, including:
[0007] Preprocessing of multi-source remote sensing data from satellite and ground sources to generate modeling auxiliary data at the 100-meter level for modeling purposes;
[0008] A virtual ground-based aerosol optical thickness monitoring network was constructed using atmospheric pollutant concentration data measured at air quality monitoring stations.
[0009] Station-scale AOD data estimated from atmospheric pollutant concentration observation stations were generated using a virtual ground-based aerosol optical thickness monitoring network.
[0010] Obtain apparent reflectance data at the 100-meter level, and then upscale the 100-meter apparent reflectance data by aggregating it through a pixel neighborhood window to obtain kilometer-level apparent reflectance data.
[0011] Based on the optimal reflectance synthesis technique, a kilometer-level surface reflectance database for the study area was constructed using kilometer-level apparent reflectance data.
[0012] Using site-scale AOD data as training labels, graph attention deep network modeling was performed with 100-meter-level modeling auxiliary data, kilometer-level surface reflectance database and 100-meter-level apparent reflectance data as input data to obtain the inversion results of 100-meter-level aerosol optical thickness.
[0013] Optionally, the satellite-ground multi-source remote sensing data can be preprocessed to generate 100-meter-level modeling aid data for modeling, including:
[0014] In multi-source remote sensing data from satellite and ground, the apparent reflectance data, snow cover data, and land use data at the 100-meter level are used to train a machine learning model of manually judged samples to perform non-reversible pixel identification in order to complete the identification processing of high-brightness pixels such as clouds / snow.
[0015] Spatial downscaling of meteorological data from multi-source satellite and ground remote sensing data;
[0016] Auxiliary data extraction and processing are performed on multi-source remote sensing data from satellite and ground.
[0017] Spatiotemporal matching processing of multi-source remote sensing data from satellite and ground;
[0018] After the above preprocessing, modeling auxiliary data at the 100-meter level is generated for modeling.
[0019] Optionally, spatial downscaling of meteorological data from multi-source satellite and ground remote sensing data can be performed, including:
[0020] A spatial downscaling model for model simulation based on machine learning was constructed using measured meteorological data from multi-source satellite and ground remote sensing data.
[0021] Model simulation spatial downscaling models are used to generate downscaled meteorological variable data at the 100-meter level. This downscaled meteorological variable data is then used to assist in the refined spatial feature mining of 100-meter AOD inversion.
[0022] Optionally, a virtual ground-based aerosol optical thickness monitoring network can be constructed using atmospheric pollutant concentration data measured at air quality monitoring stations, including:
[0023] Centered on air quality monitoring stations, measured data from an aerosol observation network within a radius of M km are obtained through spatiotemporal matching processing to generate a dataset of atmospheric pollutant concentrations.
[0024] Randomly select N% of the dataset after spatiotemporal matching as the validation set;
[0025] A virtual ground aerosol optical thickness monitoring network was constructed using measured data from a centralized aerosol observation network.
[0026] Cross-validation of the estimation accuracy of the virtual ground-based aerosol optical thickness monitoring network was performed using the validation set.
[0027] When the preset conditions are met, the virtual foundation aerosol optical thickness monitoring network is considered complete.
[0028] Optionally, acquire 100-meter-level apparent reflectance data, and upscale the 100-meter-level apparent reflectance data by aggregating it through a pixel neighborhood window to obtain kilometer-level apparent reflectance data, including:
[0029] To obtain apparent reflectance data at the 100-meter level, the apparent reflectance within the preset percentage quantile range in the corresponding band of the 100-meter level apparent reflectance data is extracted using an M*M neighborhood window combined with a pixel neighborhood window aggregation method.
[0030] After averaging the apparent reflectance within the preset percentage quantile range, kilometer-scale apparent reflectance data is obtained.
[0031] Optionally, a kilometer-scale surface reflectance database for the study area can be constructed using kilometer-scale apparent reflectance data based on optimal reflectance synthesis technology, including:
[0032] By performing spectral response function integration and least squares fitting on typical ground features, the spectral conversion coefficients between the two satellite sensor bands are obtained.
[0033] The optimal reflectance is obtained from kilometer-level apparent reflectance data based on the optimal reflectance synthesis technology.
[0034] Median filtering is used to convolve the optimal reflectance to obtain the pixel-by-pixel quarterly synthetic surface reflectance, which is then converted using spectral conversion coefficients to obtain a kilometer-level surface reflectance library for high-resolution sensors.
[0035] Optionally, using site-scale AOD data as training labels, and taking 100-meter-scale modeling auxiliary data, a kilometer-scale surface reflectance database, and 100-meter-scale apparent reflectance data as input data, a graph attention deep network model is performed to obtain the inversion results of 100-meter-scale aerosol optical thickness, including:
[0036] Using a kilometer-level surface reflectance database and a hundred-meter-level apparent reflectance data as model inputs, feature fusion is performed through a multi-head self-attention mechanism module to establish a physical mapping relationship between kilometer-level apparent reflectance and atmospheric state characteristics.
[0037] A progressive upsampling layer is designed to perform feature expansion and deconvolution. The input is 100-meter-level apparent reflectance data as a guide to spatially downscale the surface state features. Then, a front feedback layer is used to extract 100-meter-level atmospheric state features.
[0038] A sliding median filter is used to extract some pixels around the site as nodes. A graph network is constructed based on the 3D coordinate information of the nodes and edge weights are defined.
[0039] Embedding 100-meter-level modeling auxiliary data characterizing atmospheric diffusion conditions and spatial structure features into node features enhances the learning effect on the spatial coordinated change law of regional AOD.
[0040] The loss objective function is constructed by the difference between the predicted output of the deep learning model and the site-scale AOD data;
[0041] Model parameters were trained using 100-meter-level modeling auxiliary data, 100-meter-level apparent reflectance data, and site-scale AOD data as training data.
[0042] AOD inversion was performed using apparent reflectance data at the 100-meter level.
[0043] The second aspect of this application provides a high-resolution aerosol optical thickness inversion system based on graph networks, including:
[0044] The first generation unit is used to preprocess multi-source remote sensing data from satellite and ground to generate modeling auxiliary data at the 100-meter level for modeling.
[0045] The first building unit is used to construct a virtual ground-based aerosol optical thickness monitoring network using atmospheric pollutant concentration data measured at air quality monitoring stations.
[0046] The second generation unit is used to generate station-scale AOD data estimated by the atmospheric pollutant concentration observation stations using a virtual ground-based aerosol optical thickness monitoring network.
[0047] The third generation unit is used to acquire apparent reflectance data at the hundred-meter level, and upscale the hundred-meter level apparent reflectance data by aggregating the pixel neighborhood window to obtain kilometer level apparent reflectance data.
[0048] The second building unit is used to construct a kilometer-level surface reflectance database for the study area based on kilometer-level apparent reflectance data using optimal reflectance synthesis technology.
[0049] The inversion unit is used to perform graph attention deep network modeling with station-scale AOD data as training labels, and with 100-meter-level modeling auxiliary data, kilometer-level surface reflectance database and 100-meter-level apparent reflectance data as input data, to obtain the inversion results of 100-meter-level aerosol optical thickness.
[0050] Optionally, the first generating unit includes:
[0051] The first processing module is used to perform non-reversible pixel recognition on the apparent reflectance data, snow cover data and land use data at the hundred-meter level in the multi-source remote sensing data of satellite and ground, by training a machine learning model of manually discriminative samples, so as to complete the high-brightness pixel recognition processing such as cloud / snow.
[0052] The second processing module is used to perform spatial downscaling of meteorological data from multi-source satellite and ground remote sensing data.
[0053] The third processing module is used to perform auxiliary data extraction and processing on multi-source remote sensing data from satellite and ground.
[0054] The fourth processing module is used for spatiotemporal matching processing of multi-source remote sensing data from satellite and ground.
[0055] The generation module is used to generate 100-meter-level modeling auxiliary data for modeling after the above preprocessing.
[0056] Optionally, the second processing module includes:
[0057] A spatial downscaling model for model simulation based on machine learning was constructed using measured meteorological data from multi-source satellite and ground remote sensing data.
[0058] Model simulation spatial downscaling models are used to generate downscaled meteorological variable data at the 100-meter level. This downscaled meteorological variable data is then used to assist in the refined spatial feature mining of 100-meter AOD inversion.
[0059] Optional, the first building block includes:
[0060] Centered on air quality monitoring stations, measured data from an aerosol observation network within a radius of M km are obtained through spatiotemporal matching processing to generate a dataset of atmospheric pollutant concentrations.
[0061] Randomly select N% of the dataset after spatiotemporal matching as the validation set;
[0062] A virtual ground aerosol optical thickness monitoring network was constructed using measured data from a centralized aerosol observation network.
[0063] Cross-validation of the estimation accuracy of the virtual ground-based aerosol optical thickness monitoring network was performed using the validation set.
[0064] When the preset conditions are met, the virtual foundation aerosol optical thickness monitoring network is considered complete.
[0065] Optionally, the third generation unit includes:
[0066] To obtain apparent reflectance data at the 100-meter level, the apparent reflectance within the preset percentage quantile range in the corresponding band of the 100-meter level apparent reflectance data is extracted using an M*M neighborhood window combined with a pixel neighborhood window aggregation method.
[0067] After averaging the apparent reflectance within the preset percentage quantile range, kilometer-scale apparent reflectance data is obtained.
[0068] Optional, the second building block includes:
[0069] By performing spectral response function integration and least squares fitting on typical ground features, the spectral conversion coefficients between the two satellite sensor bands are obtained.
[0070] The optimal reflectance is obtained from kilometer-level apparent reflectance data based on the optimal reflectance synthesis technology.
[0071] Median filtering is used to convolve the optimal reflectance to obtain the pixel-by-pixel quarterly synthetic surface reflectance, which is then converted using spectral conversion coefficients to obtain a kilometer-level surface reflectance library for high-resolution sensors.
[0072] Optional inversion units include:
[0073] Using a kilometer-level surface reflectance database and a hundred-meter-level apparent reflectance data as model inputs, feature fusion is performed through a multi-head self-attention mechanism module to establish a physical mapping relationship between kilometer-level apparent reflectance and atmospheric state characteristics.
[0074] A progressive upsampling layer is designed to perform feature expansion and deconvolution. The input is 100-meter-level apparent reflectance data as a guide to spatially downscale the surface state features. Then, a front feedback layer is used to extract 100-meter-level atmospheric state features.
[0075] A sliding median filter is used to extract some pixels around the site as nodes. A graph network is constructed based on the 3D coordinate information of the nodes and edge weights are defined.
[0076] Embedding 100-meter-level modeling auxiliary data characterizing atmospheric diffusion conditions and spatial structure features into node features enhances the learning effect on the spatial coordinated change law of regional AOD.
[0077] The loss objective function is constructed by the difference between the predicted output of the deep learning model and the site-scale AOD data;
[0078] Model parameters were trained using 100-meter-level modeling auxiliary data, 100-meter-level apparent reflectance data, and site-scale AOD data as training data.
[0079] AOD inversion was performed using apparent reflectance data at the 100-meter level.
[0080] A third aspect of this application provides a high-resolution aerosol optical thickness inversion system based on graph networks, comprising:
[0081] Processor, memory, input / output units, and bus;
[0082] The processor is connected to memory, input / output units, and a bus;
[0083] The memory stores a program, which the processor calls to execute, such as the first aspect and any optional inversion method of the first aspect.
[0084] The fourth aspect of this application provides a high-resolution aerosol optical thickness inversion system based on graph networks, wherein the inversion system stores a program that, when executed on the inversion system, performs an inversion method as described in the first aspect and any optional inversion method of the first aspect.
[0085] As can be seen from the above technical solutions, the embodiments of this application have the following advantages:
[0086] In this application, multi-source remote sensing data from both satellite and ground sources are first preprocessed to generate 100-meter-level modeling auxiliary data, which serves as auxiliary variables in the modeling process. Next, measured atmospheric pollutant concentration data from air quality monitoring stations are acquired, and this data is used to construct a virtual ground-based aerosol optical thickness monitoring network. Then, the virtual ground-based aerosol optical thickness monitoring network is used to generate station-scale AOD data estimated for the atmospheric pollutant concentration observation stations. This station-scale AOD data is used as the ground truth label for subsequent modeling, overcoming the limitation of insufficient training samples for deep learning models in the aerosol observation network station area. Next, 100-meter-level apparent reflectance data is acquired and upscaled using pixel neighborhood window aggregation to obtain kilometer-level apparent reflectance data. Finally, a kilometer-level surface reflectance database for the study area is constructed based on the optimal reflectance synthesis technique using the kilometer-level apparent reflectance data. Finally, using station-scale AOD data as training labels, and taking 100-meter-level modeling auxiliary data, a kilometer-level surface reflectance database, and 100-meter-level apparent reflectance data as input data, a graph attention depth network was used to model the aerosol optical thickness at the 100-meter level, yielding the inversion results. This application not only overcomes the spatial heterogeneity problem of high-resolution observation data during the inversion process, but also fully mines neighborhood pixel information, while taking into account the spatial similarity and heterogeneity constraints of regional aerosol distribution, thus achieving the AOD inversion task from a high-resolution satellite. Attached Figure Description
[0087] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0088] Figure 1 This is a schematic diagram of an embodiment of the high-resolution aerosol optical thickness inversion method based on graph networks in this application;
[0089] Figure 2 This is a schematic diagram of an embodiment of the first stage of the high-resolution aerosol optical thickness inversion method based on graph networks in this application;
[0090] Figure 3 This is a schematic diagram of an embodiment of the second stage of the high-resolution aerosol optical thickness inversion method based on graph networks in this application;
[0091] Figure 4 This is a schematic diagram of an embodiment of the third stage of the high-resolution aerosol optical thickness inversion method based on graph networks in this application;
[0092] Figure 5 This is a schematic diagram of an embodiment of the high-resolution aerosol optical thickness inversion system based on graph networks of this application;
[0093] Figure 6 This is a schematic diagram of another embodiment of the high-resolution aerosol optical thickness inversion system based on graph networks of this application;
[0094] Figure 7 This is a schematic diagram of an embodiment of the high-resolution aerosol optical thickness inversion system based on graph networks of this application;
[0095] Figure 8 This is a schematic diagram of an embodiment of the accuracy verification results of the inversion data in this application;
[0096] Figure 9 This is a schematic diagram of an embodiment of the spatial distribution map of the inversion results of this application. Detailed Implementation
[0097] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.
[0098] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.
[0099] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.
[0100] As used in this application specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if detected [the described condition or event]" may be interpreted, depending on the context, as meaning "once determined," "in response to determination," "once detected [the described condition or event]," or "in response to detection [the described condition or event]."
[0101] Furthermore, in the description of this application and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.
[0102] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.
[0103] In existing technologies, traditional aerosol optical thickness (AOD) inversion methods are represented by the dark target method and the Deep Blue algorithm. The core of these algorithms is the simulation of atmospheric radiative transfer processes, and accurate land-atmosphere decoupling is key to improving the quality of AOD inversion. Therefore, to accurately solve the ill-conditioned problems in AOD inversion, the estimation of surface reflectance needs to rely on relatively stringent assumptions and necessary prior information.
[0104] To simplify the complex process of AOD physical inversion, technicians often use machine learning methods for quantitative aerosol parameter inversion. These methods typically use long-term, high-precision aerosol station measurements as the model training target, fitting the complex nonlinear relationship between AOD and apparent / surface reflectance. In these methods, the quality of the training data directly affects the model accuracy; results for areas lacking real-world ground-based observations are highly uncertain. Furthermore, these inversion methods usually perform AOD inversion pixel-by-pixel independent calculations and perform appropriate spatial aggregation operations on the inversion results. Therefore, they cannot fully exploit the rich information of neighboring pixels during the inversion process and struggle to account for the spatial similarity and heterogeneity constraints of regional aerosol distribution.
[0105] Secondly, the spatial heterogeneity of high-resolution satellite observation data is a significant issue. Traditional physical inversion methods are more sensitive to noise in the input data, which increases the uncertainty of the inversion results. Furthermore, traditional physical inversion methods lack sufficient flexibility and adaptability when processing high-resolution data, making it difficult to account for spatial details in the inversion results. Additionally, processing large-scale sample points may also present high computational complexity and time consumption. While machine learning methods guided by aerosol station observation data can overcome these problems to some extent, there is currently a lack of effective technical methods to achieve AOD inversion tasks from high-resolution satellites.
[0106] Based on this, this application discloses a high-resolution aerosol optical thickness inversion method and related apparatus, which is used to overcome the spatial heterogeneity problem of high-resolution observation data in the inversion process, and fully mine the neighboring pixel information in the inversion process, taking into account the spatial similarity and heterogeneity constraints of regional aerosol distribution, so as to realize the AOD inversion task of high-resolution satellite.
[0107] The technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.
[0108] The method described in this application can be applied to servers, devices, terminals, or other devices with logical processing capabilities; therefore, this application does not limit its application. For ease of description, the following description uses a terminal as the executing entity.
[0109] Please see Figure 1 This application provides an embodiment of a high-resolution aerosol optical thickness inversion method based on graph networks, comprising:
[0110] 101. Preprocess the satellite and ground multi-source remote sensing data to generate 100-meter-level modeling auxiliary data for modeling.
[0111] In this embodiment, the satellite-ground multi-source remote sensing data is first preprocessed. The satellite-ground multi-source remote sensing data includes, but is not limited to, WFV observation data from the Gaofen-1 satellite wide-field camera, ground-based measured atmospheric pollutant concentration data, meteorological variable data, snow cover data, surface elevation data, vegetation index data (calculated in the WFV band), and land use data.
[0112] After preprocessing the multi-source remote sensing data from satellite and ground, modeling auxiliary data at the 100-meter level is generated for modeling.
[0113] 102. Construct a virtual ground-based aerosol optical thickness monitoring network using atmospheric pollutant concentration data measured at air quality monitoring stations;
[0114] 103. Using a virtual ground-based aerosol optical thickness monitoring network, station-scale AOD data estimated by atmospheric pollutant concentration observation stations are generated;
[0115] Next, the terminal acquires measured atmospheric pollutant concentration data from air quality monitoring stations around the target location. This data is then used to construct a virtual ground-based aerosol optical thickness monitoring network. The terminal then utilizes this virtual network to generate estimated station-scale AOD data for the atmospheric pollutant concentration observation stations. This step addresses the limitation of insufficient training samples for deep learning models in the aerosol observation network's station areas.
[0116] 104. Obtain apparent reflectance data at the hundred-meter level, and then upscale the hundred-meter apparent reflectance data by aggregating it through a pixel neighborhood window to obtain kilometer-level apparent reflectance data.
[0117] 105. Based on the optimal reflectance synthesis technique, a kilometer-level surface reflectance database for the study area was constructed using kilometer-level apparent reflectance data;
[0118] After the terminal generates the estimated station-scale AOD data for atmospheric pollutant concentration observation stations using the virtual ground-based aerosol optical thickness monitoring network, it can obtain 100-meter-scale apparent reflectance data. Then, the 100-meter-scale apparent reflectance data is upscaled by aggregating pixel neighborhood windows to obtain kilometer-scale apparent reflectance data. Finally, based on the optimal reflectance synthesis technique, a kilometer-scale surface reflectance database for the study area is constructed using the kilometer-scale apparent reflectance data. For details on constructing the kilometer-scale apparent reflectance data and the kilometer-scale surface reflectance database, please refer to the subsequent instructions.
[0119] 106. Using site-scale AOD data as training labels, graph attention deep network modeling is performed with 100-meter-level modeling auxiliary data, kilometer-level surface reflectance database and 100-meter-level apparent reflectance data as input data to obtain the inversion results of 100-meter-level aerosol optical thickness.
[0120] Finally, the terminal needs to use the data obtained in the aforementioned steps as input data for modeling. Specifically, the terminal needs to use site-scale AOD data as training labels (real labels), which are used to construct the loss function. Using 100-meter-level modeling auxiliary data, a kilometer-level surface reflectance database, and 100-meter-level apparent reflectance data as input data, a graph attention deep network model is performed to obtain the inversion results of 100-meter-level aerosol optical thickness.
[0121] In this embodiment, multi-source remote sensing data from both satellite and ground sources is first preprocessed to generate 100-meter-level modeling auxiliary data, which serves as an auxiliary variable in the modeling process. Next, measured atmospheric pollutant concentration data from air quality monitoring stations are acquired, and this data is used to construct a virtual ground-based aerosol optical thickness monitoring network. Then, the virtual ground-based aerosol optical thickness monitoring network is used to generate station-scale AOD data estimated by the atmospheric pollutant concentration observation stations. This station-scale AOD data is used as the ground truth label for subsequent modeling, overcoming the limitation of insufficient training samples for deep learning models in the aerosol observation network station area. Next, 100-meter-level apparent reflectance data is acquired and upscaled using pixel neighborhood window aggregation to obtain kilometer-level apparent reflectance data. Finally, a kilometer-level surface reflectance database for the study area is constructed based on the optimal reflectance synthesis technique using the kilometer-level apparent reflectance data. Finally, using station-scale AOD data as training labels, and taking 100-meter-level modeling auxiliary data, a kilometer-level surface reflectance database, and 100-meter-level apparent reflectance data as input data, a graph attention depth network was used to model the aerosol optical thickness at the 100-meter level, yielding the inversion results. This application not only overcomes the spatial heterogeneity problem of high-resolution observation data during the inversion process, but also fully mines neighborhood pixel information, while taking into account the spatial similarity and heterogeneity constraints of regional aerosol distribution, thus achieving the AOD inversion task from a high-resolution satellite.
[0122] Please see Figure 2 , Figure 3 and Figure 4 This application provides another embodiment of a high-resolution aerosol optical thickness inversion method based on graph networks, comprising:
[0123] 201. For the apparent reflectance data, snow cover data and land use data at the hundred-meter level in the multi-source remote sensing data of satellite and ground, non-reversible pixel identification is performed by training a machine learning model of manually discriminative samples in order to complete the identification processing of high-brightness pixels such as clouds / snow.
[0124] 202. Construct a spatial downscaling model for model simulation based on machine learning using measured meteorological data from multi-source satellite and ground remote sensing data;
[0125] 203. Use the model simulation spatial downscaling model to generate downscaled meteorological variable data at the 100-meter level. The downscaled meteorological variable data at the 100-meter level is used to assist in the refined spatial feature mining of 100-meter level AOD inversion.
[0126] 204. Perform auxiliary data extraction and processing on multi-source remote sensing data from satellite and ground;
[0127] 205. Perform spatiotemporal matching processing on multi-source remote sensing data from satellite and ground;
[0128] 206. After the above preprocessing, modeling auxiliary data at the 100-meter level is generated for modeling.
[0129] In this embodiment, the multi-source remote sensing data from both satellite and ground sources is preprocessed, including but not limited to 100-meter-level apparent reflectance data, high-brightness pixel identification processing for clouds / snow, spatial downscaling of meteorological data, auxiliary data extraction processing, and spatiotemporal matching processing.
[0130] Due to the lack of infrared bands for snow and ice pixel detection (taking WFV wide-field cameras as an example), non-reversible pixel recognition is achieved by combining snow cover data and land use data and training a machine learning model of manually judged samples.
[0131] Specifically, in this embodiment, the machine learning model establishes a classification model that can be used to determine whether a pixel is invertible by using prior statistical relationships constructed from the input dataset, thus removing pixels that are not invertible due to reasons such as cloud cover. The machine learning model is as follows:
[0132] Retr 1 / 0 =f c (bands,STD,VZ,SZ,RA,SA,SI,GEO)
[0133] Among them, Retr 1 / 0 This indicates whether a pixel is a discriminant for AOD (Atlas Occurrence Displacement) pixels; bands represents satellite-observed apparent reflectance; STD represents the standard deviation of satellite apparent reflectance within a 3×3 spatial neighborhood; VZ, SZ, and RA represent the observed zenith angle, solar zenith angle, and relative azimuth angle, respectively; SA represents the scattering angle; SI represents the snow cover product; GEO represents geographic feature variables primarily based on land use and surface elevation; f c This represents a machine learning model with fitting properties, used for data classification purposes. The model is trained with parameters from manually labeled samples.
[0134] Secondly, in this embodiment, meteorological data includes, but is not limited to, temperature, relative humidity, air pressure, wind speed, and wind direction data. The terminal constructs a machine learning-based model simulation spatial downscaling model using measured meteorological data from multi-source satellite-ground remote sensing data. Specifically, the terminal uses the model simulation spatial downscaling model to generate downscaled meteorological variable data at the 100-meter level. This downscaled meteorological variable data is used to assist in the refined spatial feature mining of the 100-meter level AOD inversion. Using measured meteorological data from national meteorological stations, a machine learning-based model simulation spatial downscaling model is constructed to obtain downscaled meteorological variable data at the 100-meter level. This downscaled meteorological variable data is used to assist in the refined spatial feature mining of the 100-meter level AOD inversion. The model is as follows:
[0135] MET d =fd (MET,NDVI,GEO)
[0136] MET d The model represents downscaled meteorological variable data at the 100-meter scale, MET represents the model-simulated meteorological variable data of the meteorological variables to be downscaled, NDVI represents the 100-meter normalized vegetation index, GEO represents geographic characteristic variables mainly composed of land use and surface elevation, and f d This represents a machine learning model with regression fitting capabilities, trained using parameters from measured meteorological data from national meteorological stations.
[0137] 207. Using air quality monitoring stations as the center, obtain measured data of aerosol observation network within a radius of M km as atmospheric pollutant concentration data through spatiotemporal matching processing to generate a dataset;
[0138] 208. Randomly select N% of the dataset after spatiotemporal matching as the validation set;
[0139] 209. Construct a virtual ground aerosol optical thickness monitoring network using measured data from a centralized aerosol observation network;
[0140] 210. Use the validation set to perform cross-validation on the estimation accuracy of the virtual ground-based aerosol optical thickness monitoring network;
[0141] 211. When the preset conditions are met, the virtual foundation aerosol optical thickness monitoring network is considered complete.
[0142] In this embodiment, the terminal constructs a virtual ground-based AOD monitoring network (virtual ground-based aerosol optical thickness monitoring network) based on measured atmospheric pollutant concentration data from air quality monitoring stations. The model is as follows:
[0143] AOD PM =f r (PM 10 PM 2.5 SO2, NO2, MET d ,NDVI,GEO)
[0144] AOD PM Represents virtual foundation site-scale AOD data, PM 10 PM 2.5 SO2 and NO2 represent the near-surface inhalable particulate matter, fine particulate matter, sulfur dioxide, and nitrogen dioxide concentrations at the corresponding stations, respectively. MET d The model represents downscaled meteorological variable data at the 100-meter level, where NDVI is the 100-meter level normalized vegetation index, and GEO represents geographic characteristic variables mainly composed of land use and surface elevation. rThis refers to a machine learning model with fitting capabilities, used to achieve data regression.
[0145] Next, the parameters of the above model are trained using spatiotemporally matched data. N% (20% as an example) of the dataset after spatiotemporal matching is randomly selected as the validation set for cross-validation of model estimation accuracy, used to verify the model's generalization ability. Spatiotemporal matching refers to the matching of measured data from an aerosol observation network (AERONET) within a radius of M km (25 km as an example), centered on an air quality monitoring station.
[0146] 212. Using a virtual ground-based aerosol optical thickness monitoring network, station-scale AOD data estimated for atmospheric pollutant concentration observation stations are generated;
[0147] Step 212 in this embodiment is similar to step 103 in the previous embodiment, and will not be repeated here.
[0148] 213. Obtain apparent reflectance data at the 100-meter level, and use an M*M neighborhood window combined with a pixel neighborhood window aggregation method to extract the apparent reflectance in the corresponding band of the 100-meter level apparent reflectance data that is within the preset percentage quantile range.
[0149] 214. After averaging the apparent reflectance within the preset percentage quantile range, the kilometer-scale apparent reflectance data is obtained by upscaling.
[0150] 215. Perform spectral response function integration and least squares fitting on typical ground features to obtain the spectral conversion coefficient between the two satellite sensor bands;
[0151] 216. Based on the optimal reflectance synthesis technology, the optimal reflectance is obtained from kilometer-level apparent reflectance data;
[0152] 217. Median filtering is used to convolve the optimal reflectance to obtain the pixel-by-pixel quarterly synthetic surface reflectance, and then the kilometer-level surface reflectance library of the high-resolution sensor is obtained by converting it through the spectral conversion coefficient.
[0153] In this embodiment, the terminal uses pixel neighborhood window aggregation to generate kilometer-level apparent reflectance data. Specifically, the hundred-meter-level apparent reflectance data (taking the GF-1WFV sensor as an example) is first upscaled using pixel neighborhood window aggregation to improve the signal-to-noise ratio of the original apparent reflectance.
[0154] The aforementioned pixel neighborhood window aggregation method uses an M*M neighborhood window combined with pixel neighborhood window aggregation to extract the apparent reflectance within a preset percentage quantile range in the corresponding band of the 100-meter-level apparent reflectance data. Specifically, the terminal can use an M*M (10 as an example) neighborhood window to extract the apparent reflectance of the corresponding band that is above a preset percentage (20% as an example) quantile and below a preset percentage (50% as an example) quantile. After averaging the extracted apparent reflectance, the kilometer-level apparent reflectance is obtained by upscaling.
[0155] Considering that the satellite sensor does not include the shortwave infrared band (GF-1WFV payload) which is transparent to aerosols, the apparent reflectance products of other satellites (taking the eight-day synthetic surface reflectance product of MOD09A1 as an example) are used to construct the kilometer-level surface reflectance library of this high-resolution sensor.
[0156] The kilometer-level surface reflectance database was constructed by: performing spectral response function integration and least squares fitting on typical ground features to obtain the spectral conversion coefficients between the two satellite sensor bands; obtaining the optimal reflectance based on the optimal reflectance synthesis technique; convolving the optimal reflectance using median filtering to obtain the pixel-by-pixel quarterly synthesized surface reflectance; and finally, obtaining the kilometer-level surface reflectance database of this high-resolution sensor through the above linear relationship transformation.
[0157] Secondly, the optimal reflectance synthesis technique involves: constructing a 500-meter monthly average multi-band surface reflectance dataset using MOD09A1 data; selecting an appropriate window for time-series reflectance calculation through long-term time-series data analysis and window variance calculation to obtain the optimal reflectance.
[0158] 218. Using the kilometer-level surface reflectance database and the hundred-meter-level apparent reflectance data as model inputs, feature fusion is performed through a multi-head self-attention mechanism module to establish a physical mapping relationship between kilometer-level apparent reflectance and atmospheric state characteristics.
[0159] 219. Design a progressive upsampling layer for feature expansion and deconvolution, using 100-meter-level apparent reflectance data as a guide to spatially downscale the surface state features, and then use a front feedback layer to extract 100-meter-level atmospheric state features.
[0160] 220. Use a sliding median filter to extract some pixels around the station as nodes, construct a graph network based on the 3D coordinate information of the nodes and define edge weights;
[0161] 221. Embed the 100-meter-level modeling auxiliary data characterizing atmospheric diffusion conditions and spatial structure features into node features to enhance the learning effect on the spatial coordinated change law of regional AOD.
[0162] 222. Construct a loss objective function by using the difference between the predicted output of the deep learning model and the site-scale AOD data;
[0163] 223. Use 100-meter-level modeling auxiliary data, 100-meter-level apparent reflectance data and station-scale AOD data as training data to train model parameters;
[0164] 224. Perform AOD inversion using 100-meter-level apparent reflectance data.
[0165] In this embodiment, a kilometer-level surface reflectance database and satellite-observed hundred-meter-level apparent reflectance (hundred-meter-level apparent reflectance data) are used as model inputs. Feature fusion is performed through a multi-head self-attention mechanism module to establish a physical mapping relationship between kilometer-level apparent reflectance and atmospheric state characteristics.
[0166] Next, the terminal design uses a progressive upsampling layer to perform feature expansion and deconvolution. The input WFV 100-meter apparent reflectance (100-meter apparent reflectance data) is used as a guide to spatially downscale the surface state features. Then, a front feedback layer is used to extract 100-meter atmospheric state features.
[0167] By using a sliding median filter to extract some pixels around the station as nodes, a graph network is constructed based on the 3D coordinate information of the nodes and edge weights are defined. At the same time, auxiliary variables (hundred-meter-level modeling auxiliary data) that characterize atmospheric diffusion conditions and spatial structure features are embedded into the node features. The purpose is to enhance the learning effect on the spatial coordinated change law of regional AOD.
[0168] The deep learning model constructs a loss objective function based on the difference between the predicted output and the true label (site-scale AOD data).
[0169] Model parameters were trained using training data (100-meter-level modeling auxiliary data, 100-meter-level apparent reflectance data, and station-scale AOD data), and AOD inversion was performed using satellite observation data.
[0170] In this embodiment, firstly, non-reversible pixel identification is performed on snow cover data and land use data from multi-source remote sensing data using a machine learning model trained on manually discriminative samples to complete the identification of high-brightness pixels such as clouds / snow. A machine learning-based model simulation spatial downscaling model is constructed using measured meteorological data from multi-source remote sensing data. The model simulation spatial downscaling model is used to generate downscaled meteorological variable data at the 100-meter level, which is used to assist in the refined spatial feature mining for 100-meter AOD inversion. Auxiliary data extraction processing is performed on the multi-source remote sensing data. Spatiotemporal matching processing is then performed on the multi-source remote sensing data. After the above preprocessing, 100-meter-level modeling auxiliary data is generated for modeling, serving as auxiliary variables in the modeling process. Next, using air quality monitoring stations as the center, measured data from an aerosol observation network within a radius of M km is obtained through spatiotemporal matching processing as atmospheric pollutant concentration data to generate a dataset. N% of the dataset after spatiotemporal matching is randomly selected as the validation set. A virtual ground-based aerosol optical thickness monitoring network was constructed using measured data from a centralized aerosol observation network. The estimation accuracy of the virtual network was cross-validated using a validation set. The virtual network was considered complete when the validation met preset conditions. Then, the virtual network was used to generate estimated station-scale AOD data for atmospheric pollutant concentration observation stations. This station-scale AOD data was used as the ground truth label for subsequent modeling, overcoming the limitation of insufficient training samples for deep learning models in the aerosol observation network station areas. Next, 100-meter-level apparent reflectance data was acquired. Using an M*M neighborhood window combined with pixel neighborhood window aggregation, the apparent reflectance within the preset percentage quantile range of the corresponding bands in the 100-meter-level apparent reflectance data was extracted. The apparent reflectance within the preset percentage quantile range was averaged to obtain upscaled kilometer-level apparent reflectance data. Finally, spectral response function integration and least-squares fitting were performed on typical ground features to obtain the spectral conversion coefficients between the two satellite sensor bands. The optimal reflectance is obtained from kilometer-level apparent reflectance data using optimal reflectance synthesis technology. Median filtering is used to convolve the optimal reflectance to obtain pixel-by-pixel quarterly synthesized surface reflectance, which is then converted using spectral conversion coefficients to obtain a kilometer-level surface reflectance library from a high-resolution sensor. Finally, the kilometer-level surface reflectance library and hundred-meter-level apparent reflectance data are used as model inputs, and a multi-head self-attention mechanism module is used for feature fusion to establish a physical mapping relationship between kilometer-level apparent reflectance and atmospheric state features. A progressive upsampling layer is designed for feature expansion and deconvolution, guided by hundred-meter-level apparent reflectance data, to spatially downscale surface state features. A feedback-before-feedback layer is then used to extract hundred-meter-level atmospheric state features.A sliding median filter is used to extract pixels around the station as nodes. A graph network is constructed based on the 3D coordinate information of the nodes, and edge weights are defined. Hundred-meter-scale modeling auxiliary data, representing atmospheric diffusion conditions and spatial structure characteristics, are embedded into the node features to enhance the learning effect on the spatially coordinated variation law of regional aerosols (AOD). A loss objective function is constructed by the difference between the predicted output of the deep learning model and the station-scale AOD data. Hundred-meter-scale modeling auxiliary data, hundred-meter-scale apparent reflectance data, and station-scale AOD data are used as training data to train the model parameters. AOD inversion is performed using the hundred-meter-scale apparent reflectance data. This application not only overcomes the spatial heterogeneity problem of high-resolution observation data during the inversion process but also fully mines the information of neighboring pixels, while taking into account the spatial similarity and heterogeneity constraints of regional aerosol distribution, thus realizing the AOD inversion task of high-resolution satellites.
[0171] Please see Figure 5 This application provides an embodiment of a high-resolution aerosol optical thickness inversion system based on graph networks, comprising:
[0172] The first generation unit 301 is used to preprocess the satellite and ground multi-source remote sensing data to generate 100-meter-level modeling auxiliary data for modeling.
[0173] The first building unit 302 is used to build a virtual ground-based aerosol optical thickness monitoring network using atmospheric pollutant concentration data measured by air quality monitoring stations;
[0174] The second generation unit 303 is used to generate station-scale AOD data estimated by the atmospheric pollutant concentration observation stations using a virtual ground-based aerosol optical thickness monitoring network.
[0175] The third generation unit 304 is used to acquire 100-meter-level apparent reflectance data, and to upscale the 100-meter-level apparent reflectance data by aggregating the pixel neighborhood window to obtain kilometer-level apparent reflectance data.
[0176] The second building unit 305 is used to construct a kilometer-level surface reflectance database of the study area based on the optimal reflectance synthesis technology using kilometer-level apparent reflectance data.
[0177] Inversion unit 306 is used to perform graph attention deep network modeling with station-scale AOD data as training labels, and with 100-meter-level modeling auxiliary data, kilometer-level surface reflectance database and 100-meter-level apparent reflectance data as input data, to obtain the inversion results of 100-meter-level aerosol optical thickness.
[0178] Please see Figure 6 This application provides another embodiment of a high-resolution aerosol optical thickness inversion system based on graph networks, comprising:
[0179] The first generation unit 401 is used to preprocess the satellite and ground multi-source remote sensing data to generate 100-meter-level modeling auxiliary data for modeling.
[0180] Optionally, the first generation unit 401 includes:
[0181] The first processing module 4011 is used to perform non-reversible pixel recognition on snow cover data and land use data in multi-source remote sensing data from satellite and ground by training a machine learning model of manually discriminative samples, so as to complete the high-brightness pixel recognition processing such as cloud / snow.
[0182] The second processing module 4012 is used to perform spatial downscaling of meteorological data on multi-source remote sensing data from satellite and ground.
[0183] Optionally, the second processing module 4012 includes:
[0184] A spatial downscaling model for model simulation based on machine learning was constructed using measured meteorological data from multi-source satellite and ground remote sensing data.
[0185] Model simulation spatial downscaling models are used to generate downscaled meteorological variable data at the 100-meter level. This downscaled meteorological variable data is then used to assist in the refined spatial feature mining of 100-meter AOD inversion.
[0186] The third processing module 4013 is used to perform auxiliary data extraction and processing on multi-source remote sensing data from satellite and ground.
[0187] The fourth processing module 4014 is used for spatiotemporal matching processing of satellite and ground multi-source remote sensing data;
[0188] The generation module 4015 is used to generate 100-meter-level modeling auxiliary data for modeling after the above preprocessing.
[0189] The first building unit 402 is used to build a virtual ground-based aerosol optical thickness monitoring network using atmospheric pollutant concentration data measured by air quality monitoring stations.
[0190] Optionally, the first building unit 402 includes:
[0191] Centered on air quality monitoring stations, measured data from an aerosol observation network within a radius of M km are obtained through spatiotemporal matching processing to generate a dataset of atmospheric pollutant concentrations.
[0192] Randomly select N% of the dataset after spatiotemporal matching as the validation set;
[0193] A virtual ground aerosol optical thickness monitoring network was constructed using measured data from a centralized aerosol observation network.
[0194] Cross-validation of the estimation accuracy of the virtual ground-based aerosol optical thickness monitoring network was performed using the validation set.
[0195] When the preset conditions are met, the virtual foundation aerosol optical thickness monitoring network is considered complete.
[0196] The second generation unit 403 is used to generate station-scale AOD data estimated by the atmospheric pollutant concentration observation stations using a virtual ground-based aerosol optical thickness monitoring network.
[0197] The third generation unit 404 is used to acquire 100-meter-level apparent reflectance data, and to upscale the 100-meter-level apparent reflectance data by aggregating the pixel neighborhood window to obtain kilometer-level apparent reflectance data.
[0198] Optionally, the third generation unit 404 includes:
[0199] To obtain apparent reflectance data at the 100-meter level, the apparent reflectance within the preset percentage quantile range in the corresponding band of the 100-meter level apparent reflectance data is extracted using an M*M neighborhood window combined with a pixel neighborhood window aggregation method.
[0200] After averaging the apparent reflectance within the preset percentage quantile range, kilometer-scale apparent reflectance data is obtained.
[0201] The second building unit 405 is used to construct a kilometer-level surface reflectance database of the study area based on kilometer-level apparent reflectance data using the optimal reflectance synthesis technology.
[0202] Optionally, the second building unit 405 includes:
[0203] By performing spectral response function integration and least squares fitting on typical ground features, the spectral conversion coefficients between the two satellite sensor bands are obtained.
[0204] The optimal reflectance is obtained from kilometer-level apparent reflectance data based on the optimal reflectance synthesis technology.
[0205] Median filtering is used to convolve the optimal reflectance to obtain the pixel-by-pixel quarterly synthetic surface reflectance, which is then converted using the spectral conversion coefficient to obtain the sensor's kilometer-level surface reflectance library.
[0206] Inversion unit 406 is used to perform graph attention deep network modeling with station-scale AOD data as training labels, and with 100-meter-level modeling auxiliary data, kilometer-level surface reflectance database and 100-meter-level apparent reflectance data as input data, to obtain the inversion results of 100-meter-level aerosol optical thickness.
[0207] Optional, inversion unit 406 includes:
[0208] Using a kilometer-level surface reflectance database and a hundred-meter-level apparent reflectance data as model inputs, feature fusion is performed through a multi-head self-attention mechanism module to establish a physical mapping relationship between kilometer-level apparent reflectance and atmospheric state characteristics.
[0209] A progressive upsampling layer is designed to perform feature expansion and deconvolution. The input is 100-meter-level apparent reflectance data as a guide to spatially downscale the surface state features. Then, a front feedback layer is used to extract 100-meter-level atmospheric state features.
[0210] A sliding median filter is used to extract some pixels around the site as nodes. A graph network is constructed based on the 3D coordinate information of the nodes and edge weights are defined.
[0211] Embedding 100-meter-level modeling auxiliary data characterizing atmospheric diffusion conditions and spatial structure features into node features enhances the learning effect on the spatial coordinated change law of regional AOD.
[0212] The loss objective function is constructed by the difference between the predicted output of the deep learning model and the site-scale AOD data;
[0213] Model parameters were trained using 100-meter-level modeling auxiliary data, 100-meter-level apparent reflectance data, and site-scale AOD data as training data.
[0214] AOD inversion was performed using apparent reflectance data at the 100-meter level.
[0215] Please see Figure 7 This application provides a high-resolution aerosol optical thickness inversion system based on graph networks, comprising:
[0216] Processor 501, memory 503, input / output unit 502 and bus 504.
[0217] The processor 501 is connected to the memory 503, the input / output unit 502, and the bus 504.
[0218] The memory 503 stores a program, and the processor 501 calls the program to execute it, such as... Figure 1 , Figure 2 , Figure 3 and Figure 4 The inversion method in [the context of the text].
[0219] This application provides a high-resolution aerosol optical thickness inversion system based on graph networks. The inversion system stores a program, which executes as follows when run on the inversion system: Figure 1 , Figure 2 , Figure 3 and Figure 4 The inversion method in [the context of the text].
[0220] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.
[0221] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection between apparatuses or units through some interfaces, and may be electrical, mechanical, or other forms.
[0222] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.
[0223] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.
[0224] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
Claims
1. A method for retrieving high-resolution aerosol optical thickness based on graph networks, characterized in that, include: Preprocessing of multi-source remote sensing data from satellite and ground sources to generate modeling auxiliary data at the 100-meter level for modeling purposes; A virtual ground-based aerosol optical thickness monitoring network was constructed using atmospheric pollutant concentration data measured at air quality monitoring stations. The virtual ground-based aerosol optical thickness monitoring network is used to generate station-scale AOD data estimated by the atmospheric pollutant concentration observation stations; Obtain apparent reflectance data at the 100-meter level, and then upscale the 100-meter apparent reflectance data by aggregating it through a pixel neighborhood window to obtain kilometer-level apparent reflectance data. Based on the optimal reflectance synthesis technique, a kilometer-level surface reflectance database for the study area was constructed using kilometer-level apparent reflectance data. Using the site-scale AOD data as training labels, the hundred-meter-level modeling auxiliary data, the kilometer-level surface reflectance database, and the hundred-meter-level apparent reflectance data are used as input data to perform graph attention depth network modeling, and the inversion results of the hundred-meter-level aerosol optical thickness are obtained. Using the site-scale AOD data as training labels, the hundred-meter-level modeling auxiliary data, the kilometer-level surface reflectance database, and the hundred-meter-level apparent reflectance data are used as input data to perform graph attention deep network modeling, obtaining the inversion results of the hundred-meter-level aerosol optical thickness, including: The kilometer-level surface reflectance database and the hundred-meter-level apparent reflectance data are used as model inputs. Feature fusion is performed through a multi-head self-attention mechanism module to establish a physical mapping relationship between kilometer-level apparent reflectance and atmospheric state characteristics. A progressive upsampling layer is designed to perform feature expansion and deconvolution. The input of the 100-meter-level apparent reflectance data is used as a guide to spatially downscale the surface state features. Then, a front feedback layer is used to extract the 100-meter-level atmospheric state features. A sliding median filter is used to extract some pixels around the site as nodes. A graph network is constructed based on the 3D coordinate information of the nodes and edge weights are defined. Embedding 100-meter-level modeling auxiliary data characterizing atmospheric diffusion conditions and spatial structure features into node features enhances the learning effect on the spatial coordinated change law of regional AOD. The loss objective function is constructed by the difference between the predicted output of the deep learning model and the site-scale AOD data; Model parameters were trained using 100-meter-level modeling auxiliary data, 100-meter-level apparent reflectance data, and site-scale AOD data as training data. AOD inversion was performed using apparent reflectance data at the 100-meter level.
2. The inversion method of claim 1, wherein, The preprocessing of multi-source remote sensing data from satellite and ground to generate 100-meter-level modeling auxiliary data for modeling includes: In multi-source remote sensing data from satellite and ground, the apparent reflectance data, snow cover data, and land use data at the 100-meter level are used to train a machine learning model of manually judged samples to perform non-reversible pixel identification in order to complete the identification processing of high-brightness pixels such as clouds / snow. The aforementioned multi-source remote sensing data from satellite and ground sources undergoes spatial downscaling processing for meteorological data. Auxiliary data extraction processing is performed on the aforementioned multi-source remote sensing data from satellite and ground. Spatiotemporal matching processing is performed on the aforementioned multi-source remote sensing data from satellite and ground. After the above preprocessing, modeling auxiliary data at the 100-meter level is generated for modeling.
3. The inversion method according to claim 2, characterized in that, The meteorological data spatial downscaling process for the aforementioned multi-source remote sensing data includes: A spatial downscaling model for model simulation based on machine learning was constructed using measured meteorological data from the aforementioned multi-source remote sensing data. The model is used to simulate a spatial downscaling model to generate downscaled meteorological variable data at the 100-meter level. This downscaled meteorological variable data is used to assist in the refined spatial feature mining of 100-meter level AOD inversion.
4. The inversion method according to claim 1, characterized in that, The virtual ground-based aerosol optical thickness monitoring network is constructed using atmospheric pollutant concentration data measured at air quality monitoring stations, including: Centered on air quality monitoring stations, measured data from an aerosol observation network within a radius of M km are obtained through spatiotemporal matching processing to generate a dataset of atmospheric pollutant concentrations. After random spatiotemporal matching, N% of the dataset is selected as the validation set; A virtual ground-based aerosol optical thickness monitoring network is constructed using measured data from the aerosol observation network in the dataset. The estimation accuracy of the virtual ground-based aerosol optical thickness monitoring network was cross-validated using the validation set. When the preset conditions are met, the virtual foundation aerosol optical thickness monitoring network is considered complete.
5. The inversion method according to claim 1, characterized in that, The process of obtaining apparent reflectance data at the 100-meter level, and upscaling the 100-meter apparent reflectance data by aggregating it through a pixel neighborhood window to obtain kilometer-level apparent reflectance data, includes: To obtain apparent reflectance data at the 100-meter level, use M The apparent reflectance of the 100-meter apparent reflectance data is extracted from the corresponding band within the preset percentage quantile range by using an M-neighborhood window combined with a pixel neighborhood window aggregation method. After averaging the apparent reflectance within the preset percentage quantile range, kilometer-scale apparent reflectance data is obtained.
6. The inversion method according to claim 1, characterized in that, The optimal reflectance synthesis technique constructs a kilometer-level surface reflectance database for the study area using kilometer-level apparent reflectance data, including: By performing spectral response function integration and least squares fitting on typical ground features, the spectral conversion coefficients between the two satellite sensor bands are obtained. The optimal reflectance is obtained from kilometer-level apparent reflectance data based on the optimal reflectance synthesis technology. Median filtering is used to convolve the optimal reflectance to obtain the pixel-by-pixel quarterly synthetic surface reflectance, which is then converted using the spectral conversion coefficient to obtain a kilometer-level surface reflectance library for high-resolution sensors.
7. A high-resolution aerosol optical thickness inversion system based on graph networks, characterized in that, include: The first generation unit is used to preprocess multi-source remote sensing data from satellite and ground to generate modeling auxiliary data at the 100-meter level for modeling. The first building unit is used to construct a virtual ground-based aerosol optical thickness monitoring network using atmospheric pollutant concentration data measured at air quality monitoring stations. The second generation unit is used to generate station-scale AOD data estimated by the atmospheric pollutant concentration observation stations using the virtual ground-based aerosol optical thickness monitoring network. The third generation unit is used to acquire apparent reflectance data at the hundred-meter level, and upscale the hundred-meter level apparent reflectance data by aggregating the pixel neighborhood window to obtain kilometer level apparent reflectance data. The second building unit is used to construct a kilometer-level surface reflectance database for the study area based on kilometer-level apparent reflectance data using optimal reflectance synthesis technology. The inversion unit is used to perform graph attention depth network modeling with the station-scale AOD data as training labels, and the hundred-meter-level modeling auxiliary data, the kilometer-level surface reflectance database and the hundred-meter-level apparent reflectance data as input data to obtain the inversion result of the hundred-meter-level aerosol optical thickness. Inversion unit, including: Using a kilometer-level surface reflectance database and a hundred-meter-level apparent reflectance data as model inputs, feature fusion is performed through a multi-head self-attention mechanism module to establish a physical mapping relationship between kilometer-level apparent reflectance and atmospheric state characteristics. A progressive upsampling layer is designed to perform feature expansion and deconvolution. The input is 100-meter-level apparent reflectance data as a guide to spatially downscale the surface state features. Then, a front feedback layer is used to extract 100-meter-level atmospheric state features. A sliding median filter is used to extract some pixels around the site as nodes. A graph network is constructed based on the 3D coordinate information of the nodes and edge weights are defined. Embedding 100-meter-level modeling auxiliary data characterizing atmospheric diffusion conditions and spatial structure features into node features enhances the learning effect on the spatial coordinated change law of regional AOD. The loss objective function is constructed by the difference between the predicted output of the deep learning model and the site-scale AOD data; Model parameters were trained using 100-meter-level modeling auxiliary data, 100-meter-level apparent reflectance data, and site-scale AOD data as training data. AOD inversion was performed using apparent reflectance data at the 100-meter level.
8. A high-resolution aerosol optical thickness inversion system based on graph networks, characterized in that, Includes processor, memory, input / output units, and bus; The processor is connected to the memory, the input / output unit, and the bus; The memory stores a program, which the processor calls to execute the inversion method as described in any one of claims 1 to 6.
9. A high-resolution aerosol optical thickness inversion system based on graph networks, characterized in that, The inversion system stores a program that, when executed on the inversion system, performs the inversion method as described in any one of claims 1 to 6.
Citation Information
Patent Citations
High-resolution atmospheric aerosol inversion method based on surface reflectance library
CN111650128A
Method and system for converting graph of SQL to text into natural language statement
CN112487020A
