High-resolution aerosol property remote sensing retrieval method based on physical guidance network

Abstract

The present application relates to a high-resolution aerosol property remote sensing inversion method based on a physically guided network, comprising obtaining satellite remote sensing data and ground aerosol observation data of a target area; generating a simulation data set containing physical laws; constructing an aerosol property collaborative inversion basic network; introducing a feature conversion module into the basic network; constructing a multi-level physical constraint loss function; using the basic network and satellite remote sensing data for inversion to generate a high-resolution aerosol property spatial distribution of the target area. The present application has the beneficial effect that it effectively improves the spatial resolution and spatiotemporal robustness of remote sensing inversion of aerosols, thereby more scientifically guiding urban environmental planning and pollution control, and has important application value.

Description

Technical Field

[0001] This invention belongs to the field of environmental monitoring technology, and in particular relates to a high-resolution aerosol property remote sensing inversion method based on physical guidance networks. Background Technology

[0002] Atmospheric aerosols have a significant impact on urban air quality, climate change, and human health. Key optical properties of aerosols, such as aerosol optical depth (AOD) and single-scatter albedo (SSA), are fundamental parameters for assessing their environmental and climatic effects. Urban areas, due to dense human activity, exhibit complex aerosol composition, rapid changes, and uneven spatial distribution; therefore, high-resolution remote sensing inversion at a fine scale is of significant scientific and applied value.

[0003] Aerosol remote sensing inversion involves complex radiative transfer physics. Traditional satellite remote sensing algorithms for aerosol property inversion often require extensive prior knowledge and parameter simplification, thus focusing primarily on large-scale regional inversions with spatial resolutions greater than 1 kilometer, which is insufficient for the fine-scale monitoring needs within cities. In recent years, driven by big data, deep learning methods have shown great potential in remote sensing inversion. However, deep learning models are sensitive to data distribution and lack physical constraints, resulting in insufficient generalization ability, especially in the complex environment of cities. Therefore, there is an urgent need for a deep learning method that can incorporate the physical laws of radiative transfer to achieve high-resolution remote sensing inversion of key optical properties of aerosols at the fine-scale urban level, in order to meet the practical needs of urban aerosol environmental management. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide a high-resolution aerosol property remote sensing inversion method based on a physical guidance network.

[0005] Firstly, a high-resolution aerosol property remote sensing inversion method based on a physical guidance network is provided, including:

[0006] Step 1: Acquire satellite remote sensing data and ground aerosol observation data of the target area, and perform preprocessing;

[0007] Step 2: Based on the atmospheric radiative transfer model, simulate the physical processes of solar radiation being affected by aerosols, and generate a simulation dataset containing physical laws.

[0008] Step 3: Construct a basic network for the joint inversion of aerosol properties, and pre-train the basic network using the simulated dataset to enable the basic network to learn the physical relationship between radiation signals and aerosol properties.

[0009] Step 4: Introduce a feature transformation module into the basic network. Based on the domain adaptive method, fine-tune the network after introducing the feature transformation module using the preprocessed ground aerosol observation data.

[0010] Step 5: Construct a multi-level physical constraint loss function;

[0011] Step 6: Use the basic network and the satellite remote sensing data to perform inversion to generate a high-resolution spatial distribution of aerosol properties in the target area.

[0012] Preferably, in step 1, the satellite remote sensing data includes multispectral surface reflectance data and atmospheric top reflectance data. The multispectral data includes coastal aerosol (0.43-0.45 μm), blue light (0.45-0.51 μm), green light (0.53-0.59 μm), red light (0.64-0.67 μm), near-infrared (0.85-0.88 μm), shortwave infrared 1 (1.57-1.65 μm), and shortwave infrared 2 (2.11-2.29 μm). The ground aerosol observation data includes aerosol optical thickness (AOD) and single-scatter albedo (SSA). The preprocessing includes removing cloud contamination pixels from the satellite remote sensing data and removing outliers and calculating the ground average observation value from the ground aerosol observation data.

[0013] Preferably, in step 2, different input parameters are set for the atmospheric radiative transfer model, and the simulation dataset is output; the atmospheric radiative transfer model is selected from 6SV or MODTRAN, and the input parameters include observation angle, aerosol composition, aerosol optical thickness (AOD), and aerosol single scattering albedo (SSA).

[0014] Preferably, in step 3, the basic network for the collaborative inversion of aerosol properties is a two-stage sequence inversion network;

[0015] The first-stage network outputs preliminary inversion results for aerosol optical thickness (AOD); the second-stage network outputs inversion results for single-scattering albedo (SSA); and the calculation formula for the basic network for co-inversion of aerosol properties is as follows:

[0016]

[0017] Where Y represents the inversion result of each stage; X represents the input of each stage. The first stage input is the observation angle, multispectral surface reflectance and atmospheric top reflectance. The second stage adds aerosol optical thickness (AOD) on top of this. W and b represent the weights and biases of each hidden layer in the sub-model. Their subscripts are the corresponding hidden layer numbers. out represents the output layer. A represents the self-attention module. R represents the residual module.

[0018] Preferably, in step 4, the feature conversion module is used to map real satellite remote sensing observation data to simulated data.

[0019] Preferably, in step 5, the formula for the multi-level physical constraint loss function Loss is:

[0020]

[0021]

[0022]

[0023]

[0024] Where α1, α2 and α3 are the weights of Huber loss LHuber, correlation loss Lcorr and physical constraint loss Lcons, respectively, δ is the Huber loss hyperparameter, x and y are the predicted value and the true value, respectively, the subscript i represents the i-th sample, the superscript represents the mean, N is the total number of samples, and upper and lower are the theoretical upper and lower bounds of aerosol properties, respectively.

[0025] As a preferred option, in step 6, the accuracy of the model is verified under different numbers of neurons in the hidden layer. The accuracy of the model is evaluated by the correlation coefficient R and the root mean square error RMSE. The model with the best performance is selected for remote sensing inversion of aerosol properties to obtain the high-resolution spatial distribution of aerosol properties in the target area.

[0026] In a second aspect, a high-resolution aerosol property remote sensing inversion system based on a physical guidance network is provided for performing any of the methods described in the first aspect, including:

[0027] The acquisition module is used to acquire satellite remote sensing data and ground aerosol observation data of the target area and perform preprocessing.

[0028] The simulation module is used to simulate the physical processes of solar radiation being affected by aerosols based on the atmospheric radiative transfer model, and generate a simulation dataset containing physical laws.

[0029] The first construction module is used to construct a basic network for the joint inversion of aerosol properties. The basic network is pre-trained using the simulated dataset so that the basic network learns the physical relationship between radiation signals and aerosol properties.

[0030] An introduction module is used to introduce a feature transformation module into the basic network. Based on the domain adaptive method, the network after introducing the feature transformation module is fine-tuned using preprocessed ground aerosol observation data.

[0031] The second building module is used to construct multi-level physical constraint loss functions;

[0032] The inversion module is used to invert the basic network and the satellite remote sensing data to generate a high-resolution spatial distribution of aerosol properties in the target area.

[0033] Thirdly, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the method as described in any of the first aspects.

[0034] Fourthly, a computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a processor, implements the method as described in any of the first aspects.

[0035] The beneficial effects of this invention are as follows: First, based on an atmospheric radiative transfer model and combined with satellite spectral response simulation, a theoretical radiation-aerosol dataset is obtained for pre-training the model, enabling the network to learn the physical laws between radiation signals and aerosol properties. Further considering the issue of significant differences between real satellite remote sensing observation data and simulated data distribution under high-resolution urban conditions due to noise and other factors, a feature conversion module is designed based on the domain adaptation concept to map real satellite remote sensing observation data to ideal simulated data. This allows the network to adapt to real-world data while maintaining physical laws. Finally, a multi-level physical constraint loss function is used to improve the robustness and physical consistency of the inversion results. Compared with existing algorithms that can only obtain kilometer-level aerosol property products, this invention combines physical knowledge and a deep learning model to obtain reliable high-resolution AOD and SSA data products, effectively improving the spatial resolution and spatiotemporal robustness of remote sensing aerosol inversion. This provides more scientific guidance for urban environmental planning and pollution control, and has significant application value. Attached Figure Description

[0036] Figure 1 A flowchart of the urban high-resolution aerosol property remote sensing inversion method provided in this embodiment of the invention;

[0037] Figure 2Scatter plot for verification of the urban high-resolution aerosol property remote sensing inversion method provided in this embodiment of the invention;

[0038] Figure 3 Spatial distribution map of the urban high-resolution aerosol property remote sensing inversion method provided in this embodiment of the invention. Detailed Implementation

[0039] The present invention will be further described below with reference to embodiments. The description of the embodiments below is only for the purpose of helping to understand the present invention. It should be noted that those skilled in the art can make several modifications to the present invention without departing from the principle of the present invention, and these improvements and modifications also fall within the protection scope of the claims of the present invention.

[0040] Example 1:

[0041] To address the problems of existing technologies, Embodiment 1 of this application provides a high-resolution aerosol property remote sensing inversion method based on a physical guidance network, such as... Figure 1 As shown, it includes:

[0042] Step 1: Acquire satellite remote sensing data and ground aerosol observation data of the target area, and perform preprocessing.

[0043] In step 1, the target area can be an urban area. The satellite remote sensing data includes multispectral surface reflectance data and atmospheric top reflectance data. The multispectral data includes coastal aerosol (0.43-0.45μm), blue light (0.45-0.51μm), green light (0.53-0.59μm), red light (0.64-0.67μm), near infrared (0.85-0.88μm), shortwave infrared 1 (1.57-1.65μm), and shortwave infrared 2 (1.57-1.65μm). 2: 2.11-2.29 μm); the ground aerosol observation data includes aerosol optical thickness (AOD) and single scattering albedo (SSA); the preprocessing includes removing cloud-polluted pixels from satellite remote sensing data, and removing outliers and calculating the ground average observation value from ground aerosol observation data, for example, calculating the ground average observation value 1 hour before and after the satellite's transit.

[0044] Step 2: Based on the atmospheric radiative transfer model, simulate the physical processes of solar radiation being affected by aerosols, and generate a simulation dataset containing physical laws.

[0045] In step 2, different input parameters are set for the atmospheric radiative transfer model, and the simulated dataset is output. The atmospheric radiative transfer model is a general atmospheric radiative transfer model, such as one selected from 6SV or MODTRAN. For example, by combining satellite multispectral radiative response parameters with actual observed surface reflections, and randomly generating observation angles, aerosol composition, aerosol optical depth (AOD), and aerosol single scattering albedo (SSA) within a reasonable range as inputs to the atmospheric radiative transfer model, an ideal simulated dataset covering as many scenarios as possible is obtained.

[0046] Step 3: Construct a basic network for collaborative inversion of aerosol properties, and pre-train the basic network using the simulated dataset to enable the basic network to learn the physical correlation between radiation signals and aerosol properties.

[0047] In step 3, the basic network for the collaborative inversion of aerosol properties is a two-stage sequence inversion network;

[0048] The first-stage network outputs preliminary inversion results of aerosol optical thickness (AOD); the second-stage network, guided by the AOD results, outputs inversion results of single-scattering albedo (SSA), forming a physically related two-step inversion mechanism. The sub-model structures of the two-stage inversion networks are identical, and the calculation formula for the basic network for the co-inversion of aerosol properties is as follows:

[0049]

[0050] Where Y represents the inversion result of each stage; X represents the input of each stage. The first stage input is the observation angle, multispectral surface reflectance and atmospheric top reflectance. The second stage adds aerosol optical thickness (AOD) on top of this. W and b represent the weights and biases of each hidden layer in the sub-model. Their subscripts are the corresponding hidden layer numbers. out represents the output layer. A represents the self-attention module. R represents the residual module.

[0051] The formula for the self-attention module is expressed as:

[0052]

[0053] Where Softmax is the normalization exponential function, c is the total number of features, and Q, K, and V are Query, Key, and Value, respectively.

[0054] The residual module formula is expressed as:

[0055]

[0056] Where LayerNorm is layer normalization, Dropout is the probability of dropping neurons, and LeakyReLU is a linear rectified unit with leakage.

[0057] Step 4: Introduce a feature transformation module into the basic network. Based on the domain adaptation method, fine-tune the network after introducing the feature transformation module using preprocessed ground aerosol observation data so that the network can adapt to real-world data while maintaining physical laws.

[0058] Step 5: Construct a multi-level physical constraint loss function to improve the robustness and physical consistency of the inversion results.

[0059] Step 6: Use the basic network and the satellite remote sensing data to perform inversion to generate a high-resolution spatial distribution of aerosol properties in the target area.

[0060] Example 2:

[0061] Based on Example 1, Example 2 of this application provides a more specific method for high-resolution aerosol property remote sensing inversion based on a physical guidance network, including:

[0062] Step 1: Acquire satellite remote sensing data and ground aerosol observation data of the target area, and perform preprocessing.

[0063] Step 2: Based on the atmospheric radiative transfer model, simulate the physical processes of solar radiation being affected by aerosols, and generate a simulation dataset containing physical laws.

[0064] Step 3: Construct a basic network for collaborative inversion of aerosol properties, and pre-train the basic network using the simulated dataset to enable the basic network to learn the physical correlation between radiation signals and aerosol properties.

[0065] Step 4: Introduce a feature transformation module into the basic network. Based on the domain adaptation method, fine-tune the network after introducing the feature transformation module using preprocessed ground aerosol observation data so that the network can adapt to real-world data while maintaining physical laws.

[0066] In step 4, the feature conversion module is used to map real satellite remote sensing observation data to simulated data. Specifically, considering the significant differences in distribution between real satellite remote sensing observation data and simulated data due to noise and other factors, the feature conversion module maps real satellite remote sensing observation data to ideal simulated data for each of the two-stage inversion sub-models, based on the domain adaptation concept. This aligns the feature distributions of the two data as much as possible, aiming to enable the network to adapt to real-world data while maintaining physical laws. Furthermore, to ensure the entire network remains as consistent as possible, the structure of the feature conversion module is the same as the basic network, the only difference being that the number of neurons in the hidden layers is twice that of the basic network, and the output dimension is the same as the hidden layer dimension of the basic network, ensuring that features are fully mined and expressed.

[0067] Step 5: Construct a multi-level physical constraint loss function to improve the robustness and physical consistency of the inversion results.

[0068] In step 5, the multi-level physical constraint loss function comprehensively considers the two-stage inversion results and the physical laws of AOD and SSA. By combining Huber loss to measure the error between the predicted value and the true value, Pearson loss to ensure overall consistency, and constraint loss to ensure that the output is within a reasonable range, its formula can be expressed as:

[0069]

[0070]

[0071]

[0072]

[0073] Where α1, α2 and α3 are the weights of Huber loss LHuber, correlation loss Lcorr and physical constraint loss Lcons, respectively, δ is the Huber loss hyperparameter, x and y are the predicted value and the true value, respectively, the subscript i represents the i-th sample, the superscript represents the mean, N is the total number of samples, and upper and lower are the theoretical upper and lower bounds of aerosol properties, respectively.

[0074] Step 6: Use the basic network and the satellite remote sensing data to perform inversion to generate a high-resolution spatial distribution of aerosol properties in the target area.

[0075] In step 6, the accuracy of the model under different numbers of neurons in the hidden layer is verified. The accuracy of the model is evaluated by the correlation coefficient R and the root mean square error RMSE. The model with the best performance is selected for remote sensing inversion of aerosol properties to obtain the high-resolution spatial distribution of aerosol properties in the target area.

[0076] The effectiveness of the present invention will be further analyzed below through specific experimental results:

[0077] First, data from January 1 to December 31, 2023, were collected for the central urban area of ​​Beijing, including 30m spatial resolution multispectral surface reflectance and atmospheric top reflectance data from Landsat, and AOD and SSA observation data from the AERONET station. The data were then preprocessed according to the method described above in this invention.

[0078] Then, according to the aforementioned method of the present invention, a simulated dataset containing implicit physical laws is obtained based on the atmospheric radiative transfer model. The constructed model is pre-trained using this simulated dataset, and then the network with the added feature transformation module is fine-tuned for domain adaptation using real Landsat observation data and AERONET observation data. The model parameters are optimized through a multi-level physical constraint loss function. For example... Figure 2 As shown, the verification results indicate that the R-squared and RMSE of the model in this invention for AOD inversion are 0.85 and 0.102, respectively, and the R-squared and RMSE of SSA are 0.54 and 0.051, respectively. Furthermore, as... Figure 3 As shown, on a typical hazy day, the inversion results indicate that Beijing's AOD (Air Discharge Occurrence) is higher than the annual average, and the SSA (Spatial Aerosol Scattering) in all four bands is also higher than the annual average. This aligns with the physical principle that hazy weather causes severe aerosol pollution, and that hygroscopicity leads to greater secondary aerosol scattering. Furthermore, the inversion results are consistent with the trends observed by AERONET and provide more detailed spatial distribution information on aerosol properties within the city, which can better promote urban environmental planning and management.

[0079] It should be noted that the parts in this embodiment that are the same as or similar to those in Embodiment 1 can be referred to each other, and will not be repeated in this application.

[0080] Example 3:

[0081] Based on Example 2, Example 3 of this application provides a high-resolution aerosol property remote sensing inversion system based on a physical guidance network, including:

[0082] The acquisition module is used to acquire satellite remote sensing data and ground aerosol observation data of the target area and perform preprocessing.

[0083] The simulation module is used to simulate the physical processes of solar radiation being affected by aerosols based on the atmospheric radiative transfer model, and generate a simulation dataset containing physical laws.

[0084] The first construction module is used to construct a basic network for the joint inversion of aerosol properties. The basic network is pre-trained using the simulated dataset so that the basic network learns the physical relationship between radiation signals and aerosol properties.

[0085] An introduction module is used to introduce a feature transformation module into the basic network. Based on the domain adaptive method, the network after introducing the feature transformation module is fine-tuned using preprocessed ground aerosol observation data.

[0086] The second building module is used to construct multi-level physical constraint loss functions;

[0087] The inversion module is used to invert the basic network and the satellite remote sensing data to generate a high-resolution spatial distribution of aerosol properties in the target area.

[0088] It should be noted that the system provided in this embodiment is the corresponding system of the method provided in embodiment 2. Therefore, the parts that are the same as or similar to those in embodiment 2 in this embodiment can be referred to each other, and will not be described again in this application.

[0089] In summary, this invention provides a high-resolution aerosol property remote sensing inversion method based on a physical guidance network, which can achieve high spatial resolution AOD and SSA remote sensing inversion within cities. The results are reasonable and reliable, and have significant application value.

Claims

1. A high-resolution aerosol property remote sensing inversion method based on a physical guidance network, characterized in that, include: Step 1: Acquire satellite remote sensing data and ground aerosol observation data of the target area, and perform preprocessing; Step 2: Based on the atmospheric radiative transfer model, simulate the physical processes of solar radiation being affected by aerosols, and generate a simulation dataset containing physical laws. Step 3: Construct a basic network for the joint inversion of aerosol properties, and pre-train the basic network using the simulated dataset to enable the basic network to learn the physical relationship between radiation signals and aerosol properties. In step 3, the basic network for the collaborative inversion of aerosol properties is a two-stage sequence inversion network; The first-stage network outputs preliminary inversion results for aerosol optical thickness (AOD); the second-stage network outputs inversion results for single-scattering albedo (SSA); and the calculation formula for the basic network for co-inversion of aerosol properties is as follows: ； Where Y represents the inversion result of each stage; X represents the input of each stage. The first stage input is the observation angle, multispectral surface reflectance and atmospheric top reflectance. The second stage adds aerosol optical thickness (AOD) on top of this. W and b represent the weights and biases of each hidden layer in the sub-model. The subscripts are the corresponding hidden layer numbers. out represents the output layer. A represents the self-attention module. R represents the residual module. Step 4: Introduce a feature transformation module into the basic network. Based on the domain adaptive method, fine-tune the network after introducing the feature transformation module using the preprocessed ground aerosol observation data. Step 5: Construct a multi-level physical constraint loss function; In step 5, the formula for the multi-level physical constraint loss function Loss is: ； ； ； ； Where α1, α2, and α3 are respectively Huber loss L Huber Correlation loss L corr and physical constraint loss L cons The weights are δ, which is the Huber loss hyperparameter, x and y are the predicted and actual values, respectively, the subscript i represents the i-th sample, the superscript represents the mean, N is the total number of samples, and upper and lower are the theoretical upper and lower bounds of the aerosol properties, respectively. Step 6: Use the basic network and the satellite remote sensing data to perform inversion to generate a high-resolution spatial distribution of aerosol properties in the target area.

2. The high-resolution aerosol property remote sensing inversion method based on a physical guidance network according to claim 1, characterized in that, In step 1, the satellite remote sensing data includes multispectral surface reflectance data and atmospheric top reflectance data. The multispectral data includes coastal aerosols, blue light, green light, red light, near-infrared, shortwave infrared 1, and shortwave infrared 2. The ground aerosol observation data includes aerosol optical thickness (AOD) and single scattering albedo (SSA). The preprocessing includes removing cloud contamination pixels from the satellite remote sensing data and removing outliers and calculating the ground average observation value from the ground aerosol observation data.

3. The high-resolution aerosol property remote sensing inversion method based on a physical guidance network according to claim 2, characterized in that, In step 2, different input parameters are set for the atmospheric radiative transfer model, and the simulation dataset is output. The atmospheric radiative transfer model is selected from 6SV or MODTRAN, and the input parameters include observation angle, aerosol composition, aerosol optical thickness (AOD), and aerosol single scattering albedo (SSA).

4. The high-resolution aerosol property remote sensing inversion method based on a physical guidance network according to claim 3, characterized in that, In step 4, the feature conversion module is used to map real satellite remote sensing observation data to simulated data.

5. The high-resolution aerosol property remote sensing inversion method based on a physical guidance network according to claim 4, characterized in that, In step 6, the accuracy of the model under different numbers of neurons in the hidden layer is verified. The accuracy of the model is evaluated by the correlation coefficient R and the root mean square error RMSE. The model with the best performance is selected for remote sensing inversion of aerosol properties to obtain the high-resolution spatial distribution of aerosol properties in the target area.

6. A high-resolution aerosol property remote sensing inversion system based on a physical guidance network, characterized in that, For performing the method according to any one of claims 1 to 5, comprising: The acquisition module is used to acquire satellite remote sensing data and ground aerosol observation data of the target area and perform preprocessing. The simulation module is used to simulate the physical processes of solar radiation being affected by aerosols based on an atmospheric radiative transfer model, and to generate a simulation dataset containing physical laws. The first construction module is used to construct a basic network for the joint inversion of aerosol properties. The basic network is pre-trained using the simulated dataset so that the basic network learns the physical relationship between radiation signals and aerosol properties. An introduction module is used to introduce a feature transformation module into the basic network. Based on the domain adaptive method, the network after introducing the feature transformation module is fine-tuned using preprocessed ground aerosol observation data. The second building module is used to construct multi-level physical constraint loss functions; The inversion module is used to invert the basic network and the satellite remote sensing data to generate a high-resolution spatial distribution of aerosol properties in the target area.

7. A computer device, characterized in that, The method includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the method as claimed in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method as described in any one of claims 1 to 5.

Images