Physical constraint-based reflectance deep learning inversion method, system and electronic device

Abstract

The application discloses a physical constraint reflectivity deep learning inversion method and system and electronic equipment, and the method comprises the following steps: constructing a deep neural network model; inputting hyperspectral remote sensing sample data into the deep neural network model; using the hyperspectral remote sensing sample data to train a transmittance product combination parameter submodel, an atmospheric intrinsic reflectivity submodel and an atmospheric hemisphere albedo submodel respectively; using a physical inversion model to train the model by a physical constraint formula; and S3, acquiring hyperspectral remote sensing data of a research area, inputting the data into the deep neural network model and predicting and outputting atmospheric layer top reflectivity. The application innovatively constructs a parameter model, uses the parameter model to predict transmittance product combination parameters, atmospheric intrinsic reflectivity and atmospheric hemisphere albedo respectively, realizes efficient and rapid prediction and output of atmospheric layer top reflectivity by using three parameters and based on a constructed physical radiation transmission equation, and has the advantages of high prediction efficiency, high accuracy and the like.

Description

Technical Field

[0001] This invention relates to the field of atmospheric correction and inversion using hyperspectral images, and more particularly to a deep learning inversion method, system, and electronic device based on physically constrained reflectance. Background Technology

[0002] With the rapid development of remote sensing technology, hyperspectral remote sensing data, due to its rich spectral information and fine band division, plays an increasingly important role in resource surveys, environmental monitoring, agricultural assessment, and atmospheric research. However, the radiation information recorded by hyperspectral sensors must pass through the atmosphere before reaching the sensor, and is therefore inevitably affected by atmospheric scattering and absorption processes, resulting in a significant difference between the original observation data and the true reflectance of ground objects. Therefore, atmospheric correction processing is required. Traditional atmospheric correction processing methods mainly eliminate atmospheric influences by solving the atmospheric radiative transfer equation, such as the MODTRAN, 6S, and FLAASH models. However, these methods have certain limitations in practical applications. They rely on accurate atmospheric parameters (such as water vapor content and aerosol optical thickness), which often need to be estimated through auxiliary data or empirical models, thus introducing additional errors. The calculation process is complex, involving a large number of iterative calculations and lookup table interpolation, resulting in low computational efficiency and difficulty in meeting the needs of large-scale data processing. Moreover, the models assume uniform atmospheric conditions, while the actual atmosphere exhibits spatiotemporal heterogeneity, leading to a decrease in correction accuracy under complex atmospheric conditions.

[0003] In recent years, deep learning methods have demonstrated strong potential in the field of atmospheric correction. Compared with traditional atmospheric correction methods, deep learning has the advantages of automatically learning the complex nonlinear mapping relationship between atmospheric influences and surface reflectance from data, reducing reliance on manual feature design; high computational efficiency, enabling rapid inference after model training, making it suitable for real-time processing of massive remote sensing data; and strong adaptability, capable of learning the changing patterns of different atmospheric conditions and sensor characteristics through training data. Therefore, how to utilize the apparent reflectance received by satellite... Using deep learning to correct and invert the true reflectance of ground objects (i.e., the reflectance of the top of the atmosphere) from hyperspectral remote sensing data is a technical problem that urgently needs to be solved. Summary of the Invention

[0004] The purpose of this invention is to provide a deep learning inversion method, system, and electronic device based on physical constraints for reflectance. The deep neural network model extracts feature data from hyperspectral remote sensing data through encoding, spatial attention, decoding, and residual processing, and inputs these data into a parameter model. The parameter model combines physical constraints to predict three parameters: transmittance product combination parameter, atmospheric intrinsic reflectance, and atmospheric hemispherical albedo. Then, the physical inversion model uses these three parameters to output the predicted atmospheric top reflectance based on the physical radiative transfer equation, thus achieving high-precision atmospheric top reflectance from hyperspectral remote sensing data. The deep neural network model has advantages such as high efficiency and strong generalization ability.

[0005] The objective of this invention is achieved through the following technical solution:

[0006] A deep learning inversion method based on physically constrained reflectivity, the method comprising:

[0007] S1. Construct a deep neural network model that combines physical constraints and radiative transfer. The deep neural network model includes a parametric model and a physical inversion model. The parametric model includes a transmittance product combined parameter sub-model, an atmospheric intrinsic reflectance sub-model, and an atmospheric hemispherical albedo sub-model.

[0008] S2. Construct hyperspectral remote sensing sample data for the study area. The hyperspectral remote sensing sample data includes the apparent reflectance received by the corresponding satellite. The hyperspectral remote sensing sample data includes transmittance product combination parameter label data, atmospheric intrinsic reflectance label data, atmospheric hemispherical albedo label data, and atmospheric top reflectance label data. The transmittance is the product of the total transmittance from the Earth's surface to the sensor and the total transmittance from the Sun to the Earth's surface. Hyperspectral remote sensing sample data is input into a deep neural network model. Sub-models for transmittance product combination parameters, atmospheric intrinsic reflectance, and atmospheric hemispherical albedo are trained using hyperspectral remote sensing sample data. The physical inversion model utilizes the transmittance product combination parameters output from the deep neural network model. Atmospheric intrinsic reflectivity Atmospheric hemispherical albedo The model is trained according to the following physical constraint formula;

[0009] , , , For model parameters, Reflectance at the top of the atmosphere;

[0010] S3. Acquire hyperspectral remote sensing data of the study area and input it into the trained deep neural network model. The deep neural network model outputs the transmittance product combination parameter, atmospheric intrinsic reflectance, and atmospheric hemispherical albedo, respectively. Then, the physical inversion model uses the transmittance product combination parameter, atmospheric intrinsic reflectance, and atmospheric hemispherical albedo to output the atmospheric top reflectance.

[0011] To better realize this invention, the deep neural network model further includes an encoder, a spatial self-attention module, a decoder, and a residual optimization module. The hyperspectral remote sensing sample data or hyperspectral remote sensing data serves as the input image for the encoder. The encoder extracts multi-scale spatial features and high-dimensional semantic information from the input image to obtain multi-scale features. The spatial self-attention module is used to obtain the spatial attention weights of the multi-scale features. The decoder is used to decode the features layer by layer. The residual optimization module uses a residual connection structure to perform residual analysis on the decoded features and the input image to improve the ability to capture local features. The residual optimization module inputs the features into the parameter model. The transmittance product combination parameter sub-model, the atmospheric intrinsic reflectance sub-model, and the atmospheric hemispherical albedo sub-model of the parameter model output the parameter results of the transmittance product combination parameter, atmospheric intrinsic reflectance, and atmospheric hemispherical albedo, respectively.

[0012] Preferably, the transmittance product combined parameter sub-model also adopts the following physical constraint formula:

[0013] ,in , To enable sensors to observe the zenith angle, Atmospheric optical thickness. This represents the total diffuse transmittance of the ascending atmosphere. For diffuse transmission capability, , , , Derived from hyperspectral remote sensing data or hyperspectral remote sensing sample data, and used as essential feature terms in deep neural network models. These are the model parameters.

[0014] Preferably, the deep neural network model further includes a global information optimization module, which introduces scaling factors and offset factors to perform scaling and offset processing of features including spatial features and high-dimensional semantic information.

[0015] Preferably, the atmospheric intrinsic reflectance label data of the hyperspectral remote sensing sample data is obtained by inversion from the hyperspectral remote sensing sample data of the study area under clear and cloudless conditions, and the atmospheric hemispherical albedo label data of the hyperspectral remote sensing sample data is obtained by measuring radiation and calculating it using a four-component radiometer under clear and cloudless conditions in the study area.

[0016] Preferably, both the hyperspectral remote sensing sample data and the hyperspectral remote sensing data include satellite hyperspectral remote sensing images; in method S2, the transmittance product combination parameter sub-model, the atmospheric intrinsic reflectance sub-model, and the atmospheric hemispherical albedo sub-model of the deep neural network model are trained on the satellite hyperspectral remote sensing images pixel by pixel and spatial topological correlation model respectively; the physical inversion model is trained on the satellite hyperspectral remote sensing images pixel by pixel and spatial topological correlation model; in method S3, the deep neural network model outputs the atmospheric hemispherical albedo corresponding to each pixel of the satellite hyperspectral remote sensing image.

[0017] Preferably, the hyperspectral remote sensing sample data and hyperspectral remote sensing data include satellite hyperspectral remote sensing images acquired by satellite, and the satellite hyperspectral remote sensing images are preprocessed as follows:

[0018] A1. Perform radiometric calibration and atmospheric correction on satellite hyperspectral remote sensing images;

[0019] A2. Remove water vapor absorption bands from satellite hyperspectral remote sensing images. The wavelength range of water vapor absorption bands includes: 1363.481nm—1447.714nm, 1801.491nm—1953.111nm, and 2475.353nm—2509.043nm.

[0020] A3. Use a sliding window to extract window band data from satellite hyperspectral remote sensing images and calculate the mean and standard deviation. Remove data that exceed the mean ± 3 standard deviations within the sliding window.

[0021] Preferably, in method S2, prediction error thresholds are set for the transmittance product combined parameter sub-model, the atmospheric intrinsic reflectance sub-model, and the atmospheric hemispherical albedo sub-model, respectively. The difference between the predicted value of the transmittance product combined parameter output by the transmittance product combined parameter sub-model and the true value of the transmittance product combined parameter label data is calculated, and the difference being less than the prediction error threshold is used as a model training constraint. Similarly, the difference between the predicted value of the atmospheric intrinsic reflectance output by the atmospheric intrinsic reflectance sub-model and the true value of the atmospheric intrinsic reflectance label data is calculated, and the difference being less than the prediction error threshold is used as a model training constraint. Finally, the difference between the predicted value of the atmospheric hemispherical albedo output by the atmospheric hemispherical albedo sub-model and the true value of the atmospheric hemispherical albedo label data is calculated, and the difference being less than the prediction error threshold is used as a model training constraint.

[0022] A deep learning-based reflectance inversion system based on physical constraints includes a deep neural network model, a hyperspectral remote sensing sample database, and a hyperspectral remote sensing data acquisition module. The deep neural network model is constructed by combining physical constraints and radiative transfer. The deep neural network model includes a parametric model and a physical inversion model. The parametric model includes a transmittance product combined parameter sub-model, an atmospheric intrinsic reflectance sub-model, and an atmospheric hemispherical albedo sub-model. The hyperspectral remote sensing sample database internally stores hyperspectral remote sensing sample data of the study area and the corresponding apparent reflectance received by satellites. The hyperspectral remote sensing sample data includes transmittance product combination parameter label data, atmospheric intrinsic reflectance label data, atmospheric hemispherical albedo label data, and atmospheric top reflectance label data. The transmittance is the product of the total transmittance from the Earth's surface to the sensor and the total transmittance from the Sun to the Earth's surface. The deep neural network model's transmittance product combination parameter sub-model, atmospheric intrinsic reflectance sub-model, and atmospheric hemispherical albedo sub-model are trained using hyperspectral remote sensing sample data from the hyperspectral remote sensing sample database. The physical inversion model uses the transmittance product combination parameters output by the deep neural network model. Atmospheric intrinsic reflectivity Atmospheric hemispherical albedo The model is trained according to the following physical constraint formula; , , , For model parameters, The hyperspectral remote sensing data acquisition module is used to acquire hyperspectral remote sensing data of the study area and input it into the trained deep neural network model. The deep neural network model outputs the transmittance product combination parameter, the intrinsic atmospheric reflectance, and the atmospheric hemispherical albedo, respectively. Then, the physical inversion model uses the transmittance product combination parameter, the intrinsic atmospheric reflectance, and the atmospheric hemispherical albedo to output the atmospheric top reflectance.

[0023] An electronic device includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to cause the at least one processor to perform the steps of the physical constraint reflectivity deep learning inversion method of the present invention.

[0024] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0025] (1) The deep neural network model of the present invention extracts the feature data of hyperspectral remote sensing data through encoding, spatial attention, decoding, residual and other processing. The parameter model is combined with physical constraints to predict three parameters: transmittance product combination parameter, atmospheric intrinsic reflectance and atmospheric hemispherical albedo. Then, the physical inversion model uses the three parameters to output the predicted atmospheric top reflectance based on the physical radiative transfer equation, thus realizing the high-precision atmospheric top reflectance through hyperspectral remote sensing data. The deep neural network model has the advantages of high efficiency and strong generalization ability.

[0026] (2) This invention innovatively constructs a parametric model, which predicts the transmittance product combination parameter, atmospheric intrinsic reflectance and atmospheric hemispherical albedo respectively. It realizes the efficient and rapid prediction of the output atmospheric top reflectance using three parameters and based on the constructed physical radiation transfer equation, and has the advantages of high prediction efficiency and high accuracy.

[0027] (3) This invention explicitly embeds the physical inversion process into a deep neural network model. It learns the complex coupling relationship between atmospheric conditions and observation geometry through feature extraction, and then outputs three important intermediate parameters in parallel. Then, it uses the three parameters to construct the physical radiation transfer equation to achieve efficient inversion of the physical radiation transfer equation. By minimizing the loss function between the predicted reflectance and the true label, the gradient information is backpropagated and the three parameters are constrained and optimized, so that the network always maintains physical consistency in the data-driven learning process. Attached Figure Description

[0028] Figure 1 This is a flowchart of the physical constraint reflectivity deep learning inversion method of the present invention;

[0029] Figure 2 This is a simplified diagram illustrating the principle of the physical constraint reflectivity deep learning inversion method of this invention.

[0030] Figure 3 This is a schematic diagram illustrating the structural principle of the deep neural network model in the embodiment;

[0031] Figure 4 The example shows a loss comparison curve between the training set, validation set, and true values ​​in hyperspectral remote sensing sample data.

[0032] Figure 5 This is a schematic diagram of the root mean square error of the training set and the validation set in the embodiment;

[0033] Figure 6 This is a diagram showing the analysis results of the predicted atmospheric top reflectivity output by the model in the embodiment.

[0034] Figure 7This is a feature map of satellite hyperspectral remote sensing images of a certain study area displayed by pixel by a deep neural network model in the example;

[0035] Figures 8-12 The example diagram shows a comparison of the spectral curves between the predicted atmospheric top reflectance, the actual atmospheric top reflectance, and the apparent reflectance received by the satellite for several study areas. Detailed Implementation

[0036] The present invention will be further described in detail below with reference to embodiments:

[0037] Example

[0038] like Figure 1 As shown, a deep learning inversion method based on physical constraint reflectivity is proposed, the method comprising:

[0039] S1. Construct a deep neural network model that combines physical constraints and radiative transfer. This deep neural network model, also known as the FFH model (Fast Fusion of Hyperspectral Radiative Features), includes a parametric model and a physical inversion model. The parametric model includes a transmittance product combined parameter sub-model, an atmospheric intrinsic reflectivity sub-model, and an atmospheric hemispherical albedo sub-model. For example... Figure 2 As shown, the deep neural network model uses hyperspectral remote sensing sample data and the corresponding apparent reflectance received by satellites. Using atmospheric top reflectance data as input and atmospheric top reflectance as predicted output, the deep neural network model learns the nonlinear characteristics of the interaction between the atmosphere and the Earth's surface. It first predicts the transmittance product combination parameter, the intrinsic atmospheric reflectance, and the atmospheric hemispherical albedo. Then, it uses the transmittance product combination parameter, the intrinsic atmospheric reflectance, and the atmospheric hemispherical albedo to predict the atmospheric top reflectance based on the physical radiative transfer equation. This invention uses three parameters—transmittance product combination parameter, intrinsic atmospheric reflectance, and atmospheric hemispherical albedo—as driving variables to ensure that the deep neural network model maintains physical consistency throughout the data-driven learning process. The surface reflectance (i.e., atmospheric top reflectance) predicted by the deep neural network model is quantitatively compared with the true value to evaluate its inversion accuracy and bias.

[0040] S2. Construct hyperspectral remote sensing sample data for the study area. This data includes satellite hyperspectral remote sensing imagery (the satellite's hyperspectral sensor covers 166 bands, including visible, near-infrared, and short-wave infrared bands, offering high spectral and spatial resolution) and other data, such as sensor-observed zenith angles. Atmospheric optical thickness Total diffuse scattering transmittance of rising atmosphere Diffuse transmission capability and the apparent reflectivity received by the corresponding satellite Hyperspectral remote sensing sample data includes transmittance product combination parameter label data, atmospheric intrinsic reflectance label data, atmospheric hemispherical albedo label data, and atmospheric top reflectance label data, etc. It is the product of the total transmittance from the Earth's surface to the sensor and the total transmittance from the Sun to the Earth's surface. , Total solar transmittance to Earth's surface. The total transmittance of the surface-sensor is [missing information]. , The atmospheric intrinsic reflectance label data of the hyperspectral remote sensing sample data was obtained by inversion from hyperspectral remote sensing sample data of the study area under clear and cloudless conditions. Under clear and cloudless conditions in the study area, the sky brightness in different directions was measured and the atmospheric scattering contribution was separated to calculate the atmosphere's own reflectivity (i.e., atmospheric intrinsic reflectance). The atmospheric hemispherical albedo label data of the hyperspectral remote sensing sample data was obtained by measuring radiation with a four-component radiometer under clear and cloudless conditions in the study area and calculating it. Specifically, the four-component radiometer was used to simultaneously measure the downward total solar radiation (shortwave radiation incident on the top of the atmosphere) and the upward atmospheric reflected radiation (shortwave radiation scattered by the atmosphere and returning to space), and the ratio of the two was the atmospheric hemispherical albedo. The hyperspectral remote sensing sample data was input into a deep neural network model. The transmittance product combined parameter sub-model, the atmospheric intrinsic reflectance sub-model, and the atmospheric hemispherical albedo sub-model were trained using the hyperspectral remote sensing sample data, respectively.

[0041] In some embodiments, the deep neural network model further includes an encoder, a spatial self-attention module, a decoder, and a residual optimization module. Hyperspectral remote sensing sample data or hyperspectral remote sensing data serves as the input image for the encoder. The encoder (Encoder Block) extracts multi-scale spatial features and high-dimensional semantic information from the input image to obtain multi-scale features. The encoder (Encoder Block) includes four levels of coding units (corresponding to...) Figure 3 (Encoder Block 1 to Encoder Block 4), such as Figure 3As shown, dilated convolutions of 2x, 4x, and 8x are used to extract multi-scale spatial features and high-dimensional semantic information from the input data, significantly enhancing the network's ability to model spatial and semantic contextual information. The Spatial Attention module is used to obtain spatial attention weights for multi-scale features. This module dynamically measures the spatial relevance of features at different scales, calculating the "attention weight" of each element in the feature on other elements. The spatial self-attention mechanism considers the spatial context of the features, enabling the model to better understand and capture the dependencies between different parts of the image. The DecoderBlock decodes the features layer by layer, ultimately restoring them to the same scale as the model's input layer. The decoder uses several convolutional layers and activation functions to compress high-dimensional information into low-dimensionality, allowing the network to learn non-linear mapping relationships, better capture the structure of complex data, and possess stronger feature extraction capabilities.

[0042] The Refinement Block employs a residual connection structure to perform residual analysis on the decoded features and input image, enhancing the ability to capture local features. On one hand, the residual connection mechanism of the Refinement Block preserves low-frequency information in the image to some extent; on the other hand, it allows the network to focus on modeling the complex nonlinear parts of the surface reflectance reconstruction process. Furthermore, the residual structure further enhances the ability to capture local features, achieving super-resolution atmospheric correction. The Refinement Block inputs features into the parameter model. The parameter model's transmittance product combination parameter sub-model, atmospheric intrinsic reflectance sub-model, and atmospheric hemispherical albedo sub-model output the parameter results of transmittance product combination parameters, atmospheric intrinsic reflectance, and atmospheric hemispherical albedo, respectively. Preferably, the deep neural network model also includes a Global Information Optimization Block (Affine Block). The Global Information Optimization Block introduces scaling and offset factors to scale and offset features including spatial features and high-dimensional semantic information. The Global Information Optimization Block introduces scaling and offset factors between channels to refine the surface radiation information features between channels. Figure 7As shown, this is a feature map of a satellite hyperspectral remote sensing image (TOA) for a certain study area, processed by a deep neural network model. The row containing TOA represents the original color image (true-color image) of the satellite hyperspectral remote sensing image. The row above containing the true SR (true top atmospheric reflectance) is the feature map of the true top atmospheric reflectance. The row above containing the SR of this invention (predicted top atmospheric reflectance by the model) is the feature map of the predicted top atmospheric reflectance. All three images—TOA, true SR, and SR of this invention—are presented in true-color. The row below containing the true SR is the feature map of the true top atmospheric reflectance (using false color, a false-color image composed of band combinations). The row below containing the SR of this invention (predicted top atmospheric reflectance by the model) is the feature map of the predicted top atmospheric reflectance (using false color).

[0043] The pixel feature map of the atmospheric top reflectance predicted by the deep neural network model is close to the true value of the atmospheric top reflectance, indicating that the deep neural network model has a good prediction effect on the atmospheric top reflectance. Figure 3 As shown, the deep neural network model introduces a spatial self-attention mechanism in the encoder stage and a super-resolution module in the decoder stage. By restoring the local feature details of the image, it enhances the accuracy and clarity of the corrected surface reflectance. The network optimizes the detail representation and texture feature modeling ability of the atmospheric correction results by constructing a global information optimization module, thereby comprehensively improving the accuracy and stability of atmospheric correction. The parameters involved in the deep neural network model are shown in Table 1 below:

[0044] Table 1. Parameter Table of Deep Neural Network Models

[0045] Module Convolution parameters (channel, k, d) quantity Encoder Block1 （192，3，1） 1 Encoder Block2 （256，3，2） 1 Encoder Block3 （320，3，4） 1 Encoder Block4 （512，3，8） 1 Spatial Attention Module （512，-，-） 1 Decoder Block 1 （320，3，1） 1 Decoder Block2 （256，3，1） 1 Decoder Block3 （192，3，1） 1 Decoder Block4 （147，3，1） 1 Residual optimization module Refinement Block (147，-，-) 1 Affine Block, a global information optimization module (147，-，-) 1

[0046] Model training parameters: During training, epochs were set to 100, batch size to 8, and the initial learning rate to 0.001, using the Adam optimizer. To prevent model non-convergence due to an excessively large learning rate in the later stages of training, CosineAnnealingLR was used for learning rate scheduling, with T_max set to 100, and the rest left at their default values.

[0047] Model loss function: The Smooth L1 regression loss function is adopted. The Smooth L1 loss is a combination of L1 and L2 losses. The L1 loss is not differentiable at 0, and the L2 loss is prone to gradient explosion when the predicted value and the target value differ greatly. The smoothL1 loss improves the shortcomings of both. Piecewise function 1 is the L2 loss and piecewise function 2 is the L1 loss. The calculation formula is shown in formula (8):

[0048] (8)

[0049] In some embodiments, the transmittance product combined parameter sub-model also employs the following physical constraint formula:

[0050] ,in , To enable sensors to observe the zenith angle, Atmospheric optical thickness. This represents the total diffuse transmittance of the ascending atmosphere. For diffuse transmission capability, , , , Derived from hyperspectral remote sensing data or hyperspectral remote sensing sample data, and used as essential feature terms in deep neural network models. These are the model parameters.

[0051] The physical inversion model of this invention utilizes the transmittance product combination parameters output by a deep neural network model. Atmospheric intrinsic reflectivity Atmospheric hemispherical albedo The model is trained according to the following physical constraint formula (i.e., the physical radiative transfer equation).

[0052] , , , For model parameters, The reflectance is the top of the atmosphere. The accuracy evaluation index of the deep neural network model in this invention is: statistical analysis of the network's performance. We calculate the root mean square error (RMSE) to determine the statistical deviation between the SR value (i.e., top of the atmosphere reflectance, abbreviated as SR) predicted by the network and the true SR value (i.e., the true top of the atmosphere reflectance), as shown in equation (9). It quantifies accuracy, is easy to interpret, measures the deviation between the predicted value and the true value, and is sensitive to outliers in the data.

[0053] (9)

[0054] In the formula, for Number of pixels in the band, for The reflectance value of pixel i in the band. This is the difference between the predicted SR value (SR stands for Apron Reflectance) output by the deep neural network model and the actual SR value. For example... Figure 4 , Figure 5As shown, in this embodiment, the constructed hyperspectral remote sensing sample data and the corresponding apparent reflectance received by the satellite are divided into training and validation datasets in an 8:2 ratio. These datasets are then input into a deep neural network model for training. The loss comparison curves between the training set, validation set, and the true value are shown below. Figure 4 As shown, the root mean square error of the training set and the validation set is as follows: Figure 5 , Figure 6 As shown. See also Figure 6 , Figure 6 The root mean square error (RMSE) is represented by the red curve. Figure 6 The coefficient of determination R² is represented by the blue curve. The root mean square error (RMSE) is 0.00006992, which is extremely small. The coefficient of determination R² is 0.993237, close to 1, indicating that the model has a good fit to explain the true values. Model predictions were performed on different study areas to obtain... Figures 8-12 A comparison chart of the predicted atmospheric top reflectance, the actual atmospheric top reflectance, and the apparent reflectance received by satellite is presented (the spectral curve of the apparent reflectance received by satellite is black, the spectral curve of the actual atmospheric top reflectance is green, and the spectral curve of the predicted atmospheric top reflectance is red). The spectral curve of the apparent reflectance received by satellite is significantly affected by the atmosphere, resulting in a noticeable deviation in reflectance. The spectral curve of the predicted atmospheric top reflectance of this invention closely matches the spectral curve of the actual atmospheric top reflectance, with a smaller error and a waveform closer to the actual reflectance. This indicates that this invention demonstrates high accuracy in eliminating atmospheric influences and restoring accurate spectral information of ground objects, exhibiting strong adaptability and accuracy. It can more effectively restore the true reflectance of remote sensing data, especially under conditions of upper atmosphere and complex ground objects, showing superior performance. Compared with the actual atmospheric top reflectance, the spectral curve of the predicted atmospheric top reflectance has a smaller error and a waveform closer to the actual reflectance, indicating that the algorithm demonstrates high accuracy in eliminating atmospheric influences and restoring accurate spectral information of ground objects, exhibiting strong adaptability and accuracy.

[0055] The journey of light from the sun to the sensor can be roughly divided into the following steps: First, some sunlight fails to reach the Earth's surface and is reflected back into space by the atmosphere. This portion of the information is considered interference and is the main source of noise. Since it hasn't reached the Earth's surface, it doesn't contain any surface information; this is generally called path radiation. The remaining solar radiation reaches the target on the Earth's surface and is reflected back. This portion of solar radiation includes both direct solar radiation and diffuse radiation from the sky. Due to atmospheric scattering, some solar radiation is transmitted to the sensor, while the rest deviates from its original direction and cannot be received by the sensor. The impact of this information varies depending on the surface conditions. Under homogeneous surface conditions, the impact is small. However, under non-homogeneous surface conditions, it can lead to environmental effects. Finally, some of the solar radiation reflected from the surface is reflected back to the surface by the atmosphere and then reflected again. This process repeats itself, but after several iterations, it becomes negligible.

[0056] In remote sensing, the received signal is the signal received by the sensor. Taking the simplest form of the Earth's surface, namely the Lambertian surface, as an example, the signal received by the sensor consists of three parts: the surface reflection signal of the observed pixel after attenuation by the atmospheric beam, the scattering signal of sunlight by the atmosphere, and the signal contribution of surrounding pixels. The surface reflection signal of the observed pixel after attenuation by the atmospheric beam can be characterized by equation (1):

[0057]

[0058] in, , Total solar transmittance to Earth's surface To enable sensors to observe the zenith angle, The direct transmittance of the atmosphere in the direction of observation. Atmospheric scattering coefficient, Solar irradiance, This refers to the reflectivity at the top of the atmosphere.

[0059] The scattering signal of sunlight by the atmosphere can be characterized by equation (2):

[0060]

[0061] in, This refers to the intrinsic reflectivity of the atmosphere, i.e., path radiation.

[0062] The signal contribution of surrounding pixels can be represented by equation (3):

[0063]

[0064] The total diffuse scattering transmittance of the ascending atmosphere; , where are the surface reflectance of the target pixel (x,y); S is the atmospheric hemispherical albedo.

[0065] Adding the surface reflection signal of the observed pixel after atmospheric beam attenuation, the atmospheric scattering signal of sunlight, and the signal contribution of surrounding pixels, we can obtain the signal received by the satellite sensor as Equation (4):

[0066]

[0067] Let the total transmittance of the surface-sensor be:

[0068] The total signal received by the sensor is:

[0069] The apparent reflectance received by the satellite is then given by equation (5):

[0070]

[0071] make for Then equation (5) becomes as follows:

[0072]

[0073] Where is the intrinsic atmospheric reflectance, and S is the atmospheric hemispherical albedo. These are the reflectances of the top of the atmosphere, Apparent reflectance received by the satellite.

[0074] In some embodiments, prediction error thresholds are set for the transmittance product combined parameter sub-model, the atmospheric intrinsic reflectance sub-model, and the atmospheric hemispherical albedo sub-model, respectively. The difference between the predicted value of the transmittance product combined parameter output by the transmittance product combined parameter sub-model and the true value of the transmittance product combined parameter label data is calculated, and the difference being less than the prediction error threshold is used as a model training constraint. Similarly, the difference between the predicted value of the atmospheric intrinsic reflectance output by the atmospheric intrinsic reflectance sub-model and the true value of the atmospheric intrinsic reflectance label data is calculated, and the difference being less than the prediction error threshold is used as a model training constraint. Finally, the difference between the predicted value of the atmospheric hemispherical albedo output by the atmospheric hemispherical albedo sub-model and the true value of the atmospheric hemispherical albedo label data is calculated, and the difference being less than the prediction error threshold is used as a model training constraint.

[0075] S3. Acquire hyperspectral remote sensing data of the study area and input it into the trained deep neural network model. The deep neural network model outputs the transmittance product combination parameter, atmospheric intrinsic reflectance, and atmospheric hemispherical albedo, respectively. Then, the physical inversion model uses the transmittance product combination parameter, atmospheric intrinsic reflectance, and atmospheric hemispherical albedo to output the atmospheric top reflectance.

[0076] In some embodiments, both the hyperspectral remote sensing sample data and the hyperspectral remote sensing data include satellite hyperspectral remote sensing images; preferably, the hyperspectral remote sensing sample data and the hyperspectral remote sensing data include satellite hyperspectral remote sensing images acquired by a satellite (the hyperspectral sensor onboard the satellite covers 166 bands, including visible light, near-infrared, and short-wave infrared bands, and has high spectral and spatial resolution), and the satellite hyperspectral remote sensing images are preprocessed as follows:

[0077] A1. Perform radiometric calibration and atmospheric correction on satellite hyperspectral remote sensing imagery. Acquire raw ZY1F imagery data (i.e., hyperspectral remote sensing data containing satellite hyperspectral remote sensing imagery), and use ENVI 5.3 to perform radiometric calibration to obtain TOA (Top of Atmosphere). Then, use tools to obtain the corresponding SR (Top of Atmosphere Reflectance) of the imagery. This method yields TOA (Top of Atmosphere, the observed top of the atmosphere) and BOA (Bottom of Atmosphere, surface reflectance / bottom of the atmosphere) image pairs.

[0078] A2. Water vapor absorption bands in satellite hyperspectral remote sensing imagery were removed. The wavelength ranges of these water vapor absorption bands include: 1363.481nm–1447.714nm, 1801.491nm–1953.111nm, and 2475.353nm–2509.043nm. To improve the accuracy of atmospheric correction and reduce the impact of water vapor absorption on surface reflectance estimation, the following water vapor absorption bands in the ZY1F satellite hyperspectral data were removed: Band 98-Band 103: wavelength range 1363.481nm–1447.714nm. Strong water vapor absorption in this range significantly reduces the quality of effective information. Band 124-Band 133: wavelength range 1801.491nm–1953.111nm. The strong water vapor absorption characteristics in this spectral band affect the accuracy of surface reflectance signals. Bands 164-166 (wavelength range 2475.353nm–2509.043nm) also exhibit significant water vapor absorption, leading to a reduced signal-to-noise ratio. Band removal reduces systematic errors in the inversion process, and the model processes only high-quality information, improving training efficiency and computational performance. It also avoids interference from noisy bands in reflectivity inversion, enhancing the reliability of the final output surface reflectivity. After removing water vapor absorption bands, the remaining bands are retained in the model's input, ensuring a full reflection of surface characteristics and accurate atmospheric correction.

[0079] A3. A sliding window is used to extract window band data from satellite hyperspectral remote sensing imagery, and the mean and standard deviation are calculated. Data exceeding the mean ± 3 standard deviations within the sliding window are discarded. Example: The original data is cropped into small square blocks using a 64×64 bounding box, and the number of valid pixels for all bands within the 64×64 box is checked. If invalid pixels or outliers exist within the box, further data quality assessment steps are performed. The mean and standard deviation of each band within the box are calculated, and invalid pixels and outliers are detected. If they exceed the mean ± 3 standard deviations, the data is discarded.

[0080] The hyperspectral remote sensing sample data were preprocessed according to the methods A1 to A3 above to obtain training data. The dataset was then divided into training and validation datasets in an 8:2 ratio. The prepared data was then input into the deep neural network model for training.

[0081] In method S2, the deep neural network model's transmittance product combination parameter sub-model, atmospheric intrinsic reflectance sub-model, and atmospheric hemispherical albedo sub-model are trained on satellite hyperspectral remote sensing images using pixel-by-pixel and spatial topological correlation methods. The physical inversion model is also trained on satellite hyperspectral remote sensing images using pixel-by-pixel and spatial topological correlation methods. In method S3, the deep neural network model outputs the atmospheric hemispherical albedo corresponding to each pixel in the satellite hyperspectral remote sensing image.

[0082] A deep learning-based reflectance inversion system based on physical constraints comprises a deep neural network model, a hyperspectral remote sensing sample database, and a hyperspectral remote sensing data acquisition module. The deep neural network model is constructed by incorporating physical constraints and radiative transfer considerations. It includes a parametric model and a physical inversion model. The parametric model comprises a transmittance product combined parameter sub-model, an atmospheric intrinsic reflectance sub-model, and an atmospheric hemispherical albedo sub-model. The hyperspectral remote sensing sample database internally stores hyperspectral remote sensing sample data of the study area and the corresponding apparent reflectance received by satellites. The hyperspectral remote sensing sample data includes transmittance product combination parameter label data, atmospheric intrinsic reflectance label data, atmospheric hemispherical albedo label data, and atmospheric top reflectance label data. This is the product of the total transmittance from the Earth's surface to the sensor and the total transmittance from the Sun to the Earth's surface. The transmittance product parameter sub-model, the atmospheric intrinsic reflectance sub-model, and the atmospheric hemispherical albedo sub-model of the deep neural network model are trained using hyperspectral remote sensing sample data from the hyperspectral remote sensing sample database. The physical inversion model uses the transmittance product parameter output from the deep neural network model. Atmospheric intrinsic reflectivity Atmospheric hemispherical albedo The model is trained according to the following physical constraint formula. , , , For model parameters, This refers to the atmospheric top reflectance. The hyperspectral remote sensing data acquisition module acquires hyperspectral remote sensing data of the study area and inputs it into a trained deep neural network model. The deep neural network model outputs the transmittance product combination parameter, the intrinsic atmospheric reflectance, and the atmospheric hemispherical albedo, respectively. Then, the physical inversion model uses the transmittance product combination parameter, the intrinsic atmospheric reflectance, and the atmospheric hemispherical albedo to output the atmospheric top reflectance.

[0083] An electronic device includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to cause the at least one processor to perform the steps of the physical constraint reflectivity deep learning inversion method of this embodiment.

[0084] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A deep learning inversion method based on physically constrained reflectivity, characterized in that: The methods include: S1. Construct a deep neural network model that combines physical constraints and radiative transfer. The deep neural network model includes a parametric model and a physical inversion model. The parametric model includes a transmittance product combined parameter sub-model, an atmospheric intrinsic reflectance sub-model, and an atmospheric hemispherical albedo sub-model. S2. Construct hyperspectral remote sensing sample data for the study area. The hyperspectral remote sensing sample data includes the apparent reflectance received by the corresponding satellite. The hyperspectral remote sensing sample data includes transmittance product combination parameter label data, atmospheric intrinsic reflectance label data, atmospheric hemispherical albedo label data, and atmospheric top reflectance label data. The transmittance product combination parameter... The transmittance is the product of the total transmittance from the Earth's surface to the sensor and the total transmittance from the Sun to the Earth's surface. Hyperspectral remote sensing sample data is input into a deep neural network model. Sub-models for transmittance product combination parameters, atmospheric intrinsic reflectance, and atmospheric hemispherical albedo are trained using hyperspectral remote sensing sample data. The physical inversion model utilizes the transmittance product combination parameters output from the deep neural network model. Atmospheric intrinsic reflectivity Atmospheric hemispherical albedo The model is trained according to the following physical constraint formula; , , , For model parameters, Reflectance at the top of the atmosphere; S3. Acquire hyperspectral remote sensing data of the study area and input it into the trained deep neural network model. The deep neural network model outputs the transmittance product combination parameter, atmospheric intrinsic reflectance, and atmospheric hemispherical albedo, respectively. Then, the physical inversion model uses the transmittance product combination parameter, atmospheric intrinsic reflectance, and atmospheric hemispherical albedo to output the atmospheric top reflectance.

2. The deep learning inversion method based on physical constraint reflectivity according to claim 1, characterized in that: The deep neural network model further includes an encoder, a spatial self-attention module, a decoder, and a residual optimization module. The hyperspectral remote sensing sample data or hyperspectral remote sensing data serves as the input image for the encoder. The encoder extracts multi-scale spatial features and high-dimensional semantic information from the input image to obtain multi-scale features. The spatial self-attention module is used to obtain the spatial attention weights of the multi-scale features. The decoder is used to decode the features layer by layer. The residual optimization module uses a residual connection structure to perform residual analysis on the decoded features and the input image to improve the ability to capture local features. The residual optimization module inputs the features into the parameter model. The transmittance product combination parameter sub-model, the atmospheric intrinsic reflectance sub-model, and the atmospheric hemispherical albedo sub-model of the parameter model output the parameter results of the transmittance product combination parameter, atmospheric intrinsic reflectance, and atmospheric hemispherical albedo, respectively.

3. The deep learning inversion method based on physically constrained reflectivity according to claim 2, characterized in that: The transmittance product combined parameter sub-model also adopts the following physical constraint formula: ,in , To enable sensors to observe the zenith angle, Atmospheric optical thickness. This represents the total diffuse transmittance of the ascending atmosphere. For diffuse transmission capability, , , , Derived from hyperspectral remote sensing data or hyperspectral remote sensing sample data, and used as essential feature terms in deep neural network models. These are the model parameters.

4. The deep learning inversion method based on physically constrained reflectivity according to claim 2 or 3, characterized in that: The deep neural network model also includes a global information optimization module, which introduces scaling and offset factors to perform scaling and offset processing on features including spatial features and high-dimensional semantic information.

5. The deep learning inversion method based on physically constrained reflectivity according to claim 1, characterized in that: The atmospheric intrinsic reflectance label data of the hyperspectral remote sensing sample data was obtained by inversion from the hyperspectral remote sensing sample data of the study area under clear and cloudless conditions, and the atmospheric hemispherical albedo label data of the hyperspectral remote sensing sample data was obtained by measuring radiation and calculating it using a four-component radiometer under clear and cloudless conditions in the study area.

6. The deep learning inversion method based on physical constraint reflectivity according to claim 1, characterized in that: Both the hyperspectral remote sensing sample data and the hyperspectral remote sensing data include satellite hyperspectral remote sensing images. In method S2, the transmittance product combination parameter sub-model, the atmospheric intrinsic reflectance sub-model, and the atmospheric hemispherical albedo sub-model of the deep neural network model are trained on the satellite hyperspectral remote sensing images on a pixel-by-pixel basis and with spatial topological correlation. The physical inversion model is trained on the satellite hyperspectral remote sensing images on a pixel-by-pixel basis and with spatial topological correlation. In method S3, the deep neural network model outputs the atmospheric hemispherical albedo corresponding to each pixel of the satellite hyperspectral remote sensing image.

7. The deep learning inversion method based on physical constraint reflectivity according to claim 1, characterized in that: The hyperspectral remote sensing sample data and hyperspectral remote sensing data include satellite hyperspectral remote sensing images acquired by satellites. The satellite hyperspectral remote sensing images are preprocessed as follows: A1. Perform radiometric calibration and atmospheric correction on satellite hyperspectral remote sensing images; A2. Remove water vapor absorption bands from satellite hyperspectral remote sensing images. The wavelength range of water vapor absorption bands includes: 1363.481nm—1447.714nm, 1801.491nm—1953.111nm, and 2475.353nm—2509.043nm. A3. Use a sliding window to extract window band data from satellite hyperspectral remote sensing images and calculate the mean and standard deviation. Remove data that exceed the mean ± 3 standard deviations within the sliding window.

8. The deep learning inversion method based on physically constrained reflectivity according to claim 1, characterized in that: In method S2, prediction error thresholds are set for the transmittance product combined parameter sub-model, the atmospheric intrinsic reflectance sub-model, and the atmospheric hemispherical albedo sub-model, respectively. The difference between the predicted value of the transmittance product combined parameter output by the transmittance product combined parameter sub-model and the true value of the transmittance product combined parameter label data is calculated, and the difference being less than the prediction error threshold is used as a model training constraint. Similarly, the difference between the predicted value of the atmospheric intrinsic reflectance output by the atmospheric intrinsic reflectance sub-model and the true value of the atmospheric intrinsic reflectance label data is calculated, and the difference being less than the prediction error threshold is used as a model training constraint. Finally, the difference between the predicted value of the atmospheric hemispherical albedo output by the atmospheric hemispherical albedo sub-model and the true value of the atmospheric hemispherical albedo label data is calculated, and the difference being less than the prediction error threshold is used as a model training constraint.

9. A deep learning inversion system based on physically constrained reflectivity, characterized in that: The system includes a deep neural network model, a hyperspectral remote sensing sample database, and a hyperspectral remote sensing data acquisition module. The deep neural network model is constructed by combining physical constraints and radiative transfer. It comprises a parametric model and a physical inversion model. The parametric model includes a transmittance product combined parameter sub-model, an atmospheric intrinsic reflectance sub-model, and an atmospheric hemispherical albedo sub-model. The hyperspectral remote sensing sample database internally stores hyperspectral remote sensing sample data of the study area and the corresponding apparent reflectance received by satellites. The hyperspectral remote sensing sample data includes transmittance product combination parameter label data, atmospheric intrinsic reflectance label data, atmospheric hemispherical albedo label data, and atmospheric top reflectance label data. The transmittance is the product of the total transmittance from the Earth's surface to the sensor and the total transmittance from the Sun to the Earth's surface. The deep neural network model's transmittance product combination parameter sub-model, atmospheric intrinsic reflectance sub-model, and atmospheric hemispherical albedo sub-model are trained using hyperspectral remote sensing sample data from the hyperspectral remote sensing sample database. The physical inversion model uses the transmittance product combination parameters output by the deep neural network model. Atmospheric intrinsic reflectivity Atmospheric hemispherical albedo The model is trained according to the following physical constraint formula; , , , For model parameters, The hyperspectral remote sensing data acquisition module is used to acquire hyperspectral remote sensing data of the study area and input it into the trained deep neural network model. The deep neural network model outputs the transmittance product combination parameter, the intrinsic atmospheric reflectance, and the atmospheric hemispherical albedo, respectively. Then, the physical inversion model uses the transmittance product combination parameter, the intrinsic atmospheric reflectance, and the atmospheric hemispherical albedo to output the atmospheric top reflectance.

10. An electronic device, characterized in that: include: At least one processor; And a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to cause the at least one processor to perform the steps of the method as described in any one of claims 1 to 8.

Images