A high-resolution atmospheric aerosol inversion method, device, system and medium

Abstract

The invention discloses a high-resolution atmospheric aerosol inversion method, device and system, which relates to the field of remote sensing information technology. The satellite images of each band in the study area are uniformly resampled to 60m, and the image is then SCL The scene classification map is masked; then the atmospheric bottom reflectance data of each band and the corresponding geometric angle information are extracted, and the data set is constructed by interpolating the site data and matching it with the satellite data in time and space; finally, a deep learning model is constructed, the data is normalized and input into the model for aerosol inversion, and high-resolution aerosol optical thickness ( Aerial Optical Depth,AOD The invention has good inversion effect and high spatial resolution, and does not require a large number of parameters and pre-built lookup tables.

Description

Technical Field

[0001] The invention relates to a high-resolution atmospheric aerosol inversion method, device, system and medium, and belongs to the field of remote sensing information technology. Background Art

[0002] Atmospheric aerosols refer to solid, liquid or solid-liquid mixed particles with an aerodynamic diameter of less than 100 μm. Aerosols act on the global climate system by scattering and absorbing radiation, playing a vital role in the Earth's energy budget. They also act as cloud condensation nuclei, affecting cloud formation and precipitation. Anthropogenic aerosols have increased rapidly since the Industrial Revolution, with fossil fuel emissions accounting for the majority. These aerosols are harmful to human health and cause respiratory, cardiovascular and allergic diseases. Therefore, in order to develop appropriate aerosol mitigation strategies, a clear understanding of their distribution is crucial, especially at finer spatial scales.

[0003] At present, the mainstream research methods in aerosol inversion are based on LUT, which requires pre-building the surface reflectance library and lookup table according to the 6s model. This method has many parameters, complex models, low efficiency and other problems. There are few direct aerosol inversions based on deep learning, and the existing deep learning-based aerosol inversions are mainly aimed at low resolution. There is a large gap in high-resolution deep learning-based aerosol inversion, which cannot meet the needs of refined air quality and atmospheric pollution monitoring. Summary of the invention

[0004] The purpose of the present invention is to provide a high-resolution atmospheric aerosol inversion method, device, system and medium, which can obtain a high-resolution aerosol optical depth (AOD) distribution map through inversion to meet the needs of small-scale aerosol research and air quality monitoring.

[0005] In order to achieve the above objectives / solve the above technical problems, the present invention is implemented by adopting the following technical solutions.

[0006] In a first aspect, the present invention provides a high-resolution atmospheric aerosol inversion method, comprising:

[0007] Obtain satellite observation characteristic data of the area to be studied;

[0008] The satellite observation characteristic data of the area to be studied is input into the pre-trained deep learning aerosol inversion model for inversion processing to obtain high-resolution aerosol optical thickness (AOD) data of the area to be studied;

[0009] The deep learning aerosol inversion model performs inversion processing including: extracting features from satellite observation feature data of the area to be studied through a convolutional neural network, and increasing feature dimensions, flattening the features after increasing dimensions, flattening them, refitting the flattened features, and calculating the weights of different features through an attention mechanism layer to obtain the relationship between features and weights, and refitting the relationship between features and weights through KAN The network performs aerosol optical thickness inversion to obtain high-resolution aerosol optical thickness data for the area to be studied.

[0010] Furthermore, the method of extracting features from the satellite observation feature data of the area to be studied through a convolutional neural network and improving the feature dimension specifically includes:

[0011] The satellite observation feature data of the area to be studied are sequentially input into several 1×3 convolutional layers for feature extraction, and a ReLU process is performed after each convolutional layer to increase the feature dimension.

[0012] Furthermore, the training method of the deep learning aerosol inversion model includes:

[0013] Obtain historical satellite observation characteristic data and corresponding aerosol optical thickness data at parameter wavelengths;

[0014] The dataset is constructed based on historical satellite observation characteristic data and corresponding aerosol optical thickness data at parameter wavelengths;

[0015] The deep learning aerosol inversion model is trained using the dataset;

[0016] During the training, a control experiment was carried out through the control variable method to set the parameters of the deep learning aerosol inversion model to obtain the optimal deep learning aerosol inversion model; the control variables included training batches, training times and learning rate.

[0017] Furthermore, the method for acquiring historical satellite observation characteristic data specifically includes:

[0018] Get historical satellite image data of the set band;

[0019] The resolution of the historical satellite image data of the set band is uniformly resampled, and the mask operation is performed on the historical satellite image data of the set band. The set bands are: B1-B8A, B11-B12;

[0020] Extracting historical satellite observation feature data required for inversion from historical satellite image data after mask operation;

[0021] The historical satellite observation characteristic data include: atmospheric bottom layer reflectivity data of a set band, solar zenith angle, solar azimuth angle, and satellite zenith angle, satellite azimuth angle data and normalized vegetation index corresponding to each set band.

[0022] Furthermore, the step of constructing a data set based on the historical satellite observation characteristic data and the aerosol optical thickness data of the corresponding parameter wavelength specifically includes the following steps:

[0023] The aerosol optical thickness data of the parameter wavelength is interpolated to the set wavelength to obtain the aerosol optical thickness of the set wavelength;

[0024] The satellite observation characteristic data and the aerosol optical thickness at the set wavelength are spatially and temporally matched.

[0025] Furthermore, interpolating the aerosol optical thickness data of the parameter wavelength to the set wavelength to obtain the aerosol optical thickness of the set wavelength specifically includes:

[0026] According to the fitting relationship between aerosol optical thickness and wavelength, the aerosol optical thickness data of the parameter wavelength is interpolated by a quadratic polynomial, and finally the aerosol optical thickness value of the set wavelength is obtained, wherein the aerosol optical thickness of the parameter wavelength is the AOD value at 440nm, 500nm, and 675nm, and the aerosol optical thickness of the set wavelength is the AOD value at 50nm;

[0027] The fitting expression is as follows:

[0028] lnτ(λ)=a0+a1lnλ+a2(lnλ) 2 ;

[0029] Where τ(λ) represents the aerosol optical thickness at wavelength λ, and a0, a1 and a2 are fitting parameters;

[0030] Accurate atmospheric calibration of the “cleanest” image requires the input of the AOD value at 550nm, but the ground aerosol detection system does not have AOD data at this wavelength. Therefore, the existing observed wavelength and AOD data are used to calculate the AOD data of the wavelength to be measured by difference, which has high numerical accuracy.

[0031] Furthermore, the space matching method comprises:

[0032] The satellite observation characteristic data are divided according to the longitude and latitude of the global automatic observation network site (AERONET) corresponding to the aerosol optical thickness data of the set wavelength;

[0033] Taking the global automatic observation network site (AERONET) as the center, the average value of the reflectance of each band of the pixel points in the surrounding area is set to represent the reflectance information of the site;

[0034] The time matching method comprises:

[0035] Match the aerosol optical thickness at the set wavelength of the global automatic observation network site before and after the satellite passes with the satellite observation characteristic data;

[0036] In a second aspect, the present invention provides a high-resolution atmospheric aerosol inversion device, comprising:

[0037] Acquisition module: used to obtain satellite observation characteristic data of the area to be studied;

[0038] Aerosol inversion module: used to input the satellite observation characteristic data of the area to be studied into the pre-trained deep learning aerosol inversion model, perform inversion processing, and obtain high-resolution aerosol optical thickness data of the area to be studied;

[0039] The deep learning aerosol inversion model performs inversion processing including: extracting features from satellite observation feature data of the area to be studied through a convolutional neural network, and increasing feature dimensions, flattening the features after increasing dimensions, flattening them, refitting the flattened features, and calculating the weights of different features through an attention mechanism layer to obtain the relationship between features and weights, and refitting the relationship between features and weights through KAN The network performs aerosol optical thickness inversion to obtain high-resolution aerosol optical thickness data for the area to be studied.

[0040] In a third aspect, the present invention provides a high-resolution atmospheric aerosol inversion system, comprising:

[0041] Memory, for storing computer programs / instructions;

[0042] A processor is used to execute the computer program / instructions to implement the steps of the above-mentioned high-resolution atmospheric aerosol inversion method.

[0043] In a fourth aspect, a computer-readable storage medium stores a computer program / instruction thereon, which, when executed by a processor, implements the above-mentioned high-resolution atmospheric aerosol inversion steps.

[0044] Compared with the prior art, the beneficial effects achieved by the present invention are as follows: the high-resolution atmospheric aerosol inversion method based on deep learning of the present invention constructs a deep learning AOD inversion model, and adds a convolutional neural network and an attention mechanism on the basis of the KAN network, thereby enhancing the feature extraction and information processing capabilities of the KAN network. The convolutional neural network is good at extracting multi-level spatial features from the input data and can better capture local information and patterns in the data, while the attention mechanism enables the model to focus on the most important parts of the input data by assigning different weights; in this way, the improved KAN network can not only learn the complex correlations between the data more fully, but also significantly improve the overall accuracy of the AOD inversion model, so that the model performs more accurately and reliably when facing complex and changeable data, thereby obtaining a more accurate high-resolution aerosol optical thickness distribution map.

[0045] The present invention can use the atmospheric bottom reflectivity data and corresponding geometric angle information of high-resolution satellite images and the real-time AOD data of AERONET (Global Automatic Observation Network) ground stations to construct a deep learning data set, and then normalize the data set, input the data set into the model for model training and model verification, and continuously adjust and optimize the model through feedback information to obtain a deep learning aerosol inversion model.

[0046] When performing aerosol inversion, the present invention does not require a large number of parameters and cumbersome construction of lookup tables. It only needs to obtain satellite reflectivity and geometric angle data and input them into the deep learning aerosol inversion model. The present invention can be applied to various types of surfaces, has fewer restrictions, and has wide application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 This is a flow chart for implementing the high-resolution atmospheric aerosol inversion method of the present invention;

[0048] Figure 2 This is the Sentinel-2 RGB color satellite image of Beijing on February 26, 2023;

[0049] Figure 3 This is the Sentinel-2 snow and ice mask result for Beijing 20230226;

[0050] Figure 4 This is the 20220127 Sentinel-2 RGB color satellite image;

[0051] Figure 5 This is the 20220127 Sentinel-2 ice and snow mask result;

[0052] Figure 6This is the Sentinel-2 RGB color satellite image of Beijing on September 14, 2023;

[0053] Figure 7 This is the cloud and water mask result of Sentinel-2 in Beijing on September 14, 2023;

[0054] Figure 8 This is the 20220805 Sentinel-2 RGB color satellite image;

[0055] Fig. 9 This is the 20220805 Sentinel-2 cloud water mask result;

[0056] Fig.10 This is a diagram of the high-resolution atmospheric aerosol inversion network structure based on deep learning;

[0057] Fig.11 Verify the AOD inversion results of Beijing station;

[0058] Fig.12 Verify the AOD inversion results of Beijing_PKU station;

[0059] Fig.13 Verify the AOD inversion results of Beijing_RADI station;

[0060] Fig.14 Verify the inversion results of Beijing_CAMS station;

[0061] Fig.15 Verify the AOD inversion results of XiangHe station;

[0062] Fig.16 Verification results of all Sentinel-2 AOD inversion results and AERONET AOD measurements. DETAILED DESCRIPTION

[0063] The technical solution of the present invention is described in detail below through the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations on the technical solution of the present invention. The embodiments of the present invention and the technical features in the embodiments may be combined with each other unless there is a conflict.

[0064] The term "and / or" is only a description of the association relationship between related objects, indicating that there can be three relationships. For example, A and / or B can mean: A exists alone, A and B exist at the same time, and B exists alone. In addition, the character " / " generally indicates that the related objects are in an "or" relationship.

[0065] Example 1

[0066] like Figure 1 An embodiment shown in the figure provides a high-resolution atmospheric aerosol inversion method, comprising the following steps:

[0067] Step 1: Obtain high-resolution satellite images of the study area and ground station AOD data:

[0068] Acquire historical satellite image data information of the study area through high-resolution satellites; in this embodiment, the multispectral image data of the Sentinel-2 high-resolution satellite is used to explore the study area;

[0069] In this embodiment, the remote sensing images used are Level-2A remote sensing images of Sentinel-2 in Beijing from 2019 to 2023;

[0070] In this embodiment, the ground real-time AOD data is obtained through the AERONET station, including the AOD data information of the five stations: Beijing, Beijing_CAMS, Beijing_PUK, Beijing_RADI, and XiangHe. At the same time, in order to reduce the difference of the data and increase the availability of the data, the Level 1.5 (cloud screening and quality control) level data is used for the inversion effect verification.

[0071] like Figures 2 to 9 As shown in step 2, set the band satellite image to be uniformly resampled and perform mask processing:

[0072] Since Sentinel-2 has three different resolutions of 10, 20, and 60 meters according to different bands, they are uniformly resampled to 60m resolution;

[0073] Since clouds, cloud shadows, water, snow and other factors have always been important interference factors in remote sensing aerosol inversion, the inversion results are easily affected by them. Removing pixels "polluted" by clouds, cloud shadows, water, snow and other factors in the satellite images of the study area can reduce the inversion error of atmospheric aerosols. Therefore, in this embodiment, according to the Sentinel-2 SCL scene classification map, these interference factors are masked to remove the influence of these interference factors.

[0074] Step 3: Extract the satellite observation feature data required for inversion from the satellite image data after the mask operation:

[0075] The reflectivity of all bands (B1-B8A, B11-B12) of each pixel and its corresponding geometric angle data information (solar zenith angle, solar azimuth angle, satellite zenith angle, satellite azimuth angle data) are extracted from the satellite remote sensing image of the mask.

[0076] Step 4: Construct a deep learning aerosol inversion dataset, which mainly includes the following steps:

[0077] Step 4.1: Accurate AOD verification data acquisition:

[0078] The high-resolution atmospheric aerosol inversion based on deep learning outputs the AOD value at 550nm, but the ground-based AERONRT station does not provide data in this band, so interpolation processing is required to convert the AOD value to a wavelength of 550nm.

[0079] In this embodiment, the AOD values ​​at 440 nm, 500 nm, and 675 nm are interpolated by a quadratic polynomial to obtain the AOD value at 550 nm. The specific fitting relationship is as follows:

[0080] lnτ(λ)=a0+a1lnλ+a2(lnλ) 2 ;

[0081] Where τ(λ) represents the AOD value at wavelength λ, a0, a1 and a2 are fitting parameters;

[0082] Step 4.2, data space matching:

[0083] Since aerosols in the atmosphere are mobile and the longitude and latitude of the station do not completely correspond to the longitude and latitude of each pixel in the satellite image, it is necessary to accurately characterize the reflectivity information of the station;

[0084] In this embodiment, the masked data is divided according to the longitude and latitude of the station, and the reflectivity information of the station is represented by the average value of the reflectivity of each band of 5*5 pixel points centered on the station.

[0085] Step 4.3, data time matching:

[0086] The ground-based AERONET station takes AOD values ​​approximately every 15 minutes during peak periods, while Sentinel-2 provides a satellite remote sensing image approximately every 2-3 days. Considering the mobility of aerosols, time matching is required.

[0087] In this embodiment, the AERONET site measurement values ​​within 30 minutes before and after the satellite transit are selected to match the satellite data; if there are multiple groups of site measurement values ​​in the interval, these values ​​are averaged to represent the AOD value at the current moment.

[0088] Step 5: Build a deep learning aerosol inversion model:

[0089] Step 5.1, data set normalization:

[0090] In this embodiment, since the reflectivity data of each band and the geometric angle data are not in the same order of magnitude, in order to improve the model accuracy and accelerate the model convergence speed, the input data needs to be normalized according to a unified rule. The normalization formula is as follows:

[0091] ;

[0092] Among them: X is the original data set, X min is the minimum value in the data set, X max is the maximum value in the data set, X scaled is the normalized data;

[0093] Step 5.2: Construct a deep learning aerosol inversion model based on the KAN network:

[0094] The KAN network is based on the Kolmogorov-Arnold representation theorem: any multivariable continuous function f can be represented as a combination of a finite number of single-variable continuous functions, that is,

[0095] ;

[0096] in , , ; represents a continuous multivariable function, represents a univariate function, represents the constant coefficient, is the input variable, and n represents the number of variables;

[0097] Unlike traditional neural networks that use fixed activation functions, KAN uses learnable activation functions at the edge of the network. This design allows each weight parameter in KAN to be replaced by a univariate function, which is usually parameterized in the form of spline functions. This provides extremely high flexibility and the ability to simulate complex functions with fewer parameters, enhancing the interpretability of the model.

[0098] In this embodiment, the input data for training the deep learning aerosol inversion model includes the reflectance of each band of B1-B8A, B11-B12, and the observation zenith angle, observation azimuth and solar zenith angle, solar azimuth, normalized difference vegetation index (NDVI), a total of 36 characteristic quantities, and the AOD value of 550nm, so as to learn the potential relationship between the data.

[0099] In this embodiment, ;

[0100] Among them: B8 represents the reflectivity of the bottom atmosphere in the B8 band, and B4 represents the reflectivity of the bottom atmosphere in the B4 band.

[0101] like Fig.10 As shown in the figure, the input data set is first convolved, passing through 4 1×3 convolutional layers, each followed by a ReLU activation function, and the dimension is increased to 1152; in this process, the convolution layer (ConvNet class) is mainly used to extract features from the input data, and convolution operations are performed on the input data through sliding convolution kernels to extract useful local features. At the same time, the convolution layer makes the model more generalizable and reduces the risk of overfitting by sharing convolution kernel parameters; after convolution, the data enters the flatten layer for flattening, flattening the data into a one-dimensional vector for use in the subsequent fully connected layer;

[0102] Then, the fully connected layer (FC) is refitted to reduce the loss of feature information; first, the input feature is transformed into an attention weight vector through a two-layer fully connected network. fc1 (the first fully connected layer) performs a linear transformation on the input data, converts the dimension to hidden_dim (64 dimensions), and then uses the tanh activation function to make the weight between [-1, 1]. Next, fc2 (the second fully connected layer) reduces the dimension to 1 to generate the attention weight. The attention weight is normalized by the softmax function to obtain the weight value of each input feature. These weights indicate the degree of influence that the model believes the feature has on the final prediction. A new feature representation is obtained by multiplying the input feature with the corresponding attention weight and then summing all the features. This weighted sum operation enables the model to pay more attention to important features that are assigned high weights and ignore the minor parts.

[0103] After feature extraction, the data enters the KAN network. The traditional weight parameters are replaced by univariate function parameters at the edge of the network. In KAN, each node summarizes the output of these functions without any nonlinear transformation:

[0104] );

[0105] Where: KAN ( ) represents the output of KAN, represents the function matrix of the Lth layer, The splines are the core of the KAN learning mechanism. They replace the traditional weight parameters commonly used in neural networks. The flexibility of splines enables them to adaptively model complex relationships in the data by adjusting their shape, thereby minimizing the approximation error and enhancing the network's ability to learn subtle patterns from high-dimensional data sets. The general formula of splines in KAN can be expressed as B-splines:

[0106] ;

[0107] in represents the spline function, represents the coefficients optimized during training, represents the B-spline basis functions defined on a grid.

[0108] Step 5.3, parameter adjustment and model establishment:

[0109] When training the deep learning aerosol inversion model, the parameters thereof need to be set, including training batches, number of training times, learning rate, etc. These parameters play a vital role in the training efficiency, convergence speed and accuracy of the model. In this embodiment, the training batches are designed to be 16, 32, 36, 64, and 72, the number of training times are designed to be 500, 800, 1000, 1200, 1500, 2000, and 3000, and the learning rate is designed to be 0.1, 0.01, 0.001, 0.0001, etc. Control experiments are designed in turn by the control variable method, and it is finally determined that the model achieves the best effect when the training batch is 36, the number of training times is 1000, and the learning rate is 0.001.

[0110] Step 6: AOD inversion based on deep learning:

[0111] The high-resolution satellite image dataset of the study area is normalized and input into the deep learning aerosol inversion model to invert the AOD value at 550nm and obtain a high-resolution aerosol optical thickness spatial distribution map of the study area.

[0112] In this embodiment, the input data for the deep learning aerosol inversion model test includes the reflectivity of each band of B1-B8A, B11-B12, and the observation zenith angle, observation azimuth and solar zenith angle, solar azimuth, and normalized difference vegetation index (NDVI), totaling 36 features; the output data is the AOD value of 550nm, and the 550nm AOD obtained by interpolation of the AERONET site is used to verify the model effect; in order to verify the inversion effect, the correlation coefficient R, root mean square error RMSE, mean absolute error MAE, and expected error EE are used to evaluate it, and the relevant calculation formulas are as follows:

[0113] ;

[0114] ;

[0115] ;

[0116] ;

[0117] in: represents the i-th AOD value predicted by the model, represents the ith AOD value of the AETRONET station.

[0118] This dataset has a total of 1352 samples, of which 80% (1081) are used for model training and the remaining 20% ​​(271) are used for model accuracy verification. The specific AOD inversion results are compared with the AOD measurement values ​​of the AERONET station. Figures 11 to 16 As shown in the figure, there are 1352 samples in this dataset, of which 80% (1081) are used for model training (training set), and the remaining 20% ​​(271) are used for model accuracy verification (test set). It can be seen that the overall inversion results have a high correlation with the AOD measurements of the AERONET station, with R reaching 0.9312, RMSE being 0.1167, and MAE being 0.0759. 77.86% of the inversion results are within the expected error EE. Among the five stations, the correlation of the Beijing_PKU station is the highest, reaching 0.9452. The correlation of the Beijing station is relatively low, but still 0.9054. These results show that the high-resolution atmospheric aerosol inversion method based on deep learning proposed in this invention can invert regional aerosols more accurately, overcoming the shortcomings of traditional inversion methods with many parameters, complex models, and low efficiency.

[0119] Example 2

[0120] This embodiment provides a high-resolution atmospheric aerosol inversion device, including:

[0121] Acquisition module: used to obtain satellite observation characteristic data of the area to be studied;

[0122] Aerosol inversion module: used to input the satellite observation characteristic data of the area to be studied into the pre-trained deep learning aerosol inversion model, perform inversion processing, and obtain a high-resolution aerosol optical thickness spatial distribution map of the area to be studied;

[0123] The deep learning aerosol inversion model performs inversion processing including: extracting features from satellite observation feature data of the area to be studied through a convolutional neural network, and increasing feature dimensions, flattening the features after increasing dimensions, flattening them, refitting the flattened features, and calculating the weights of different features through an attention mechanism layer to obtain the relationship between features and weights, and refitting the relationship between features and weights through KAN The network performs aerosol optical thickness inversion to obtain high-resolution aerosol optical thickness data for the area to be studied.

[0124] Example 3

[0125] This embodiment provides a high-resolution atmospheric aerosol inversion system, including:

[0126] Memory, for storing computer programs / instructions;

[0127] A processor is used to execute the computer program / instructions to implement the steps of the above-mentioned high-resolution atmospheric aerosol inversion method.

[0128] Example 4

[0129] This embodiment provides a computer-readable storage medium, including: a computer program / instruction stored thereon, and when the computer program / instruction is executed by a processor, the above-mentioned steps of high-resolution atmospheric aerosol inversion are implemented.

[0130] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0131] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0132] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0133] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0134] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the enlightenment of the present invention, ordinary technicians in this field can also make many forms without departing from the scope of protection of the purpose of the present invention and the claims, which all fall within the protection of the present invention.

Claims

1. A high-resolution atmospheric aerosol inversion method, characterized in that: include: Obtain satellite observation characteristic data of the area to be studied; The satellite observation characteristic data of the area to be studied is input into the pre-trained deep learning aerosol inversion model for inversion processing to obtain high-resolution aerosol optical thickness data of the area to be studied; Among them: the inversion processing of the deep learning aerosol inversion model includes: extracting features from the satellite observation feature data of the area to be studied through a convolutional neural network, and improving the feature dimension, flattening the features after the dimension improvement, flattening them, re-fitting the flattened features, and calculating the weights of different features through the attention mechanism layer to obtain the relationship between features and weights, and inverting the aerosol optical thickness through the KAN network using the relationship between features and weights to obtain high-resolution aerosol optical thickness data of the area to be studied.

2. The high-resolution atmospheric aerosol inversion method according to claim 1, characterized in that: The method of extracting features from satellite observation feature data of the area to be studied through a convolutional neural network and improving feature dimensions specifically includes: The satellite observation feature data of the area to be studied are sequentially input into several 1×3 convolutional layers for feature extraction, and a ReLU process is performed after each convolutional layer to increase the feature dimension.

3. The high-resolution atmospheric aerosol inversion method according to claim 1, characterized in that: The training method of the deep learning aerosol inversion model includes: Obtain historical satellite observation characteristic data and corresponding aerosol optical thickness data at parameter wavelengths; The dataset is constructed based on historical satellite observation characteristic data and corresponding aerosol optical thickness data at parameter wavelengths; The deep learning aerosol inversion model is trained using the dataset; During the training, a control experiment was carried out through the control variable method to set the parameters of the deep learning aerosol inversion model to obtain the optimal deep learning aerosol inversion model; the control variables included training batches, training times and learning rate.

4. The high-resolution atmospheric aerosol inversion method according to claim 3, characterized in that: The method for acquiring historical satellite observation feature data specifically includes: Get historical satellite image data of the set band; The resolution of the historical satellite image data of the set band is uniformly resampled, and the mask operation is performed on the historical satellite image data of the set band. The set bands are: B1-B8A, B11-B12; Extracting historical satellite observation feature data required for inversion from historical satellite image data after mask operation; The historical satellite observation characteristic data include: atmospheric bottom layer reflectivity data of a set band, solar zenith angle, solar azimuth angle, and satellite zenith angle, satellite azimuth angle data and normalized vegetation index corresponding to each set band.

5. The high-resolution atmospheric aerosol inversion method according to claim 3, characterized in that: The step of constructing a data set based on historical satellite observation characteristic data and aerosol optical thickness data of corresponding parameter wavelengths specifically includes the following steps: The aerosol optical thickness data of the parameter wavelength is interpolated to the set wavelength to obtain the aerosol optical thickness of the set wavelength; Spatial and temporal matching is performed on historical satellite observation characteristic data and aerosol optical thickness at a set wavelength.

6. The high-resolution atmospheric aerosol inversion method according to claim 5, characterized in that: The step of interpolating the aerosol optical thickness data of the parameter wavelength to the set wavelength to obtain the aerosol optical thickness of the set wavelength specifically includes: According to the fitting relationship between aerosol optical thickness and wavelength, the aerosol optical thickness data of the parameter wavelength is interpolated by a quadratic polynomial, and finally the aerosol optical thickness value of the set wavelength is obtained; The fitting expression is as follows: lnτ(λ)=a0+a1lnλ+a2(lnλ) 2 ; where τ(λ) represents the aerosol optical thickness at wavelength λ, and a0, a1, and a2 are fitting parameters.

7. The high-resolution atmospheric aerosol inversion method according to claim 5, characterized in that: The space matching method comprises: The satellite observation characteristic data are divided according to the longitude and latitude of the global automatic observation site corresponding to the aerosol optical thickness data of the set wavelength; Taking the global automatic observation site as the center, the average reflectance of each band of the pixel points in the surrounding area is set to represent the reflectance information of the site; The time matching method comprises: The aerosol optical depth at a set wavelength at a global automatic observation site before and after the satellite passes is matched with the satellite observation characteristic data.

8. A high-resolution atmospheric aerosol inversion device, characterized in that: include: Acquisition module: used to obtain satellite observation characteristic data of the area to be studied; Aerosol inversion module: used to input the satellite observation characteristic data of the area to be studied into the pre-trained deep learning aerosol inversion model, perform inversion processing, and obtain high-resolution aerosol optical thickness data of the area to be studied; The deep learning aerosol inversion model performs inversion processing including: extracting features from satellite observation feature data of the area to be studied through a convolutional neural network, and increasing feature dimensions, flattening the features after increasing dimensions, flattening them, refitting the flattened features, and calculating the weights of different features through an attention mechanism layer to obtain the relationship between features and weights, and refitting the relationship between features and weights through KAN The network performs aerosol optical thickness inversion to obtain high-resolution aerosol optical thickness data for the area to be studied.

9. A high-resolution atmospheric aerosol inversion system, characterized in that: include: Memory, for storing computer programs / instructions; A processor, configured to execute the computer program / instructions to implement the steps of the high-resolution atmospheric aerosol inversion method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: A computer program / instruction is stored thereon, and when the computer program / instruction is executed by a processor, the steps of the high-resolution atmospheric aerosol inversion method described in any one of claims 1 to 7 are implemented.

Images