Infrared image adaptive atmospheric correction method and device based on radiation transfer model

Abstract

The application discloses an infrared image adaptive atmospheric correction method and device based on a radiation transmission model, and belongs to the technical field of remote sensing image radiation correction. The method first extracts the spatial range, imaging time and observation geometry parameters of an image scene, constructs a scene-based atmospheric profile grid by distance inverse ratio weighted interpolation based on a reanalysis atmospheric profile, and combines a digital elevation model and a view zenith angle to correct and apply terrain and observation geometry constraints to generate control node input parameters. Then, the atmospheric radiation transmission model is called in batches to obtain the hyperspectral results of each control node, and the transmittance and path radiance parameters of the infrared target band are extracted by spectral response function convolution. Finally, the control node parameters are reprojected to the original image coordinates, the four-neighbor-domain bilinear interpolation and nine-neighbor-domain weighted interpolation results are adaptively fused for each pixel, the full-resolution atmospheric parameter field is recovered, and the observed radiance at the ground is inversed.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image radiometric correction technology, and in particular to an adaptive atmospheric correction method and device for infrared images based on a radiative transfer model. This method enables quantitative remote sensing applications such as refined atmospheric correction of infrared remote sensing images and radiance inversion at the Earth's surface, in conjunction with actual image data. Background Technology

[0002] Atmospheric correction of satellite remote sensing imagery has always been a crucial step in applications such as quantitative inversion, multi-temporal comparison, change detection, and surface parameter inversion. Its purpose is to minimize the impact of atmospheric absorption, scattering, and path radiation on sensor observation signals, thereby improving the accuracy, comparability, and quantitative application capabilities of remote sensing data. For infrared remote sensing imagery, the accuracy of atmospheric correction results directly affects the precision of subsequent inversions of surface radiance, surface reflectance, and related parameters.

[0003] In existing technologies, atmospheric correction of infrared remote sensing imagery typically combines standard atmospheric models, regionally unified atmospheric parameters, or utilizes global reanalysis data (such as fifth-generation ECMWF atmospheric reanalysis data, ERA5) to construct atmospheric profiles, and then uses radiative transfer models to obtain parameters such as transmittance and path radiance. ERA5 data has been widely used in remote sensing atmospheric correction due to its advantages of temporal continuity, wide spatial coverage, and ease of acquisition. However, the spatial resolution of ERA5 atmospheric profile data is typically only 0.25°. For medium-to-high spatial resolution satellite imagery, a large number of pixels within a large imaging area often share the same set or a few sets of atmospheric profile parameters. While this type of method is simple to implement and computationally efficient, it usually assumes that the atmospheric state within the area is spatially approximately uniform, making it difficult to accurately reflect the spatial differences in atmospheric components such as water vapor and ozone within a local area. It also fails to adequately consider the impact of complex terrain and observational geometry on atmospheric transport processes, resulting in deviations between key correction parameters such as transmittance and path radiance and the actual situation, thus affecting the stability and reliability of surface radiative inversion results. Another type of approach focuses more on correction accuracy, typically improving the representation of real-world scenes through finer-scale atmospheric parameter modeling, terrain constraints, and observational geometry corrections. However, these methods often require determining atmospheric profiles, terrain conditions, and observational geometry parameters for smaller scale units or even per pixel, and then using radiative transfer models for solution. For wide-swath, high-resolution infrared remote sensing images, the number of pixels typically reaches millions or even higher. Directly performing pixel-by-pixel atmospheric correction would not only be computationally intensive but also difficult to meet practical needs in terms of data organization, model running efficiency, and operational applications. Therefore, how to effectively reduce the computational complexity of large-scale image processing while ensuring the accuracy of infrared image atmospheric correction has become a pressing technical problem in this field.

[0004] Therefore, it is necessary to propose an infrared image atmospheric correction method that can both improve the physical consistency of atmospheric correction by utilizing reanalysis atmospheric data and radiative transfer models, and reduce the overall computational complexity through scene-based parameter construction and pixel-level adaptive processing, so as to balance the efficiency requirements of large-scale image processing and the accuracy requirements of atmospheric correction in complex scenes. Based on this, this invention proposes an adaptive atmospheric correction method for infrared remote sensing images based on a radiative transfer model. Through scene-based atmospheric parameter construction, radiative transfer modeling, and pixel-level adaptive fusion, it improves the accuracy of atmospheric correction for large-scale images while ensuring computational efficiency. Summary of the Invention

[0005] This invention aims to propose an adaptive atmospheric correction method and apparatus for infrared images based on a radiative transfer model. This addresses the problems of low spatial resolution of atmospheric profiles, insufficient representation of local topography and apparent geometric differences, and excessively low computational efficiency for pixel-by-pixel radiative transfer calculations in existing infrared image atmospheric correction methods. While maintaining the constraints of the physical mechanism of radiative transfer, this method combines scene-based atmospheric parameter construction with pixel-level adaptive fusion. This allows atmospheric correction to reflect spatially non-uniform atmospheric conditions while avoiding the enormous computational burden of directly running the radiative transfer model pixel-by-pixel on the entire image, thus balancing correction accuracy and processing efficiency.

[0006] This invention adopts the following technical solution: an adaptive atmospheric correction method for infrared images based on a radiative transfer model, comprising the following steps:

[0007] S1: Acquire remote sensing images, reanalyze atmospheric profiles, digital elevation models and spectral response functions, and extract the spatial extent of the scene, imaging time and observation geometric parameters as a unified scene description;

[0008] S2: Based on the discrete grid points of the reanalysis atmospheric profile, a scene-based atmospheric profile grid is constructed by inverse distance weighted interpolation, and the surface elevation of each grid point is corrected according to the digital elevation model. At the same time, the apparent zenith angle of each grid point relative to the scene center is calculated according to the observation geometric parameters to obtain the control node input parameters for solving the radiative transfer problem.

[0009] S3: Write the input parameters of each control node into the standard input file of the atmospheric radiative transfer model in batches, call the atmospheric radiative transfer model to perform batch calculations, and extract the hyperspectral radiative transfer results of each control node from the standard output file.

[0010] S4: Convolve the hyperspectral radiative transfer result with the spectral response function to extract the atmospheric transmittance and path radiative parameters of each control node in each infrared target band, forming a control node parameter grid.

[0011] S5: Reproject the control node parameter grid onto the original image coordinate system. For each original image pixel, construct a four-neighbor bilinear interpolation result and a nine-neighbor distance-weighted interpolation result respectively. Calculate the adaptive fusion weight based on the variance and gradient information of the local window. Fuse the two interpolation results into a full-resolution atmospheric transmittance field and a path radiation field. Then, based on the radiative transfer inversion relationship between entrance pupil radiance, path radiation, and transmittance, calculate the observed radiance at the ground surface pixel by pixel.

[0012] An adaptive atmospheric correction device for infrared images based on a radiative transfer model, comprising:

[0013] The parameter acquisition module acquires remote sensing images, reanalysis atmospheric profiles, digital elevation models and spectral response functions, and extracts the spatial extent of the scene, imaging time and observation geometric parameters as a unified scene description.

[0014] The parameter solving module constructs a scene-based atmospheric profile grid based on the discrete grid points of the reanalysis atmospheric profile through inverse distance weighted interpolation, corrects the surface elevation of each grid point according to the digital elevation model, and calculates the apparent zenith angle of each grid point relative to the scene center according to the observation geometric parameters to obtain the control node input parameters for solving radiative transfer.

[0015] The result extraction module writes the input parameters of each control node into the standard input file of the atmospheric radiative transfer model in batches, calls the atmospheric radiative transfer model to perform batch calculations, and extracts the hyperspectral radiative transfer results of each control node from the standard output file.

[0016] The parameter grid extraction module convolves the hyperspectral radiative transfer result with the spectral response function to extract the atmospheric transmittance and path radiative parameters of each control node in each infrared target band, forming a control node parameter grid.

[0017] The radiance calculation module reprojects the control node parameter grid onto the original image coordinate system. For each original image pixel, it constructs a four-neighbor bilinear interpolation result and a nine-neighbor distance-weighted interpolation result, respectively. Based on the variance and gradient information of the local window, it calculates an adaptive fusion weight and fuses the two interpolation results into a full-resolution atmospheric transmittance field and a path radiation field. Then, based on the radiative transfer inversion relationship between entrance pupil radiance, path radiation, and transmittance, it calculates the observed radiance at the ground surface pixel by pixel.

[0018] An electronic device includes: one or more processors; and a memory for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the method.

[0019] A computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, cause the processor to implement the method described thereon.

[0020] The present invention has the following beneficial effects:

[0021] This invention addresses the problem in existing infrared image atmospheric correction where large areas often share the same set of atmospheric parameters, making it difficult to reflect local atmospheric differences. By combining scene spatial location, terrain undulation, and observation geometry, it constructs key physical quantities such as transmittance and path radiance to more accurately reflect actual distributions. This process prevents different regions from simply applying the same atmospheric conditions. Especially in mid-wave infrared imaging, which is highly sensitive to water vapor changes, it reduces correction deviations caused by atmospheric parameter mismatches, making subsequent recovery of surface radiance from entrance pupil radiance more closely resemble reality.

[0022] This invention employs a processing method that combines control node calculation with whole-scene restoration. While preserving the physical meaning of the radiative transfer model, it avoids the high computational load caused by directly solving for radiative transfer pixel by pixel in the whole-scene image. In other words, this invention does not abandon physical modeling, but maintains model constraints in the stage of calculating physical quantities such as transmittance and path radiation, and then restores the model in the spatial unfolding stage, thus balancing the accuracy and efficiency required for large-scale infrared image processing.

[0023] In the parameter recovery stage, this invention employs a pixel-by-pixel adaptive fusion method, fusing the two candidate interpolation results according to an adaptive window. Through this method, physical quantities such as transmittance and path radiance no longer exhibit obvious blocky abrupt changes in space, but rather achieve a more natural and continuous transition with terrain, viewing angle, and local variation features. Therefore, the resulting surface radiance observation results are spatially smoother and better meet the actual needs of atmospheric correction for infrared images. Attached Figure Description

[0024] Figure 1 The flowchart shows the adaptive atmospheric correction method for infrared images based on the radiative transfer model.

[0025] Figure 2 Here is a practical example of this method in Wide Swath Imager (WSI) images: Figure 2 In the figure, (a) represents transmittance. Figure 2 (b) in the text represents path radiation. Figure 2 (c) in the figure represents the DEM elevation. Figure 2 (d) in the image represents the radiance inversion image observed at the Earth's surface. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other. To achieve the above objectives, this invention adopts the following technical solution.

[0027] This invention proposes an adaptive atmospheric correction method for infrared images based on a radiative transfer model, comprising the following steps:

[0028] S1: Read remote sensing image data as input data and extract key parameters such as time, location, and viewing angle.

[0029] First, the remote sensing image data to be corrected is read, including the external auxiliary data corresponding to the infrared image, including reanalysis atmospheric profile data, image metadata, digital elevation model (DEM) data, band spectral response function, and reference projection image. Basic information such as the latitude and longitude of the image center, the four corner coordinates, the imaging time, the central observation zenith angle, and the range of the study area are extracted from these data.

[0030] Function: To establish a unified scene description required for subsequent atmospheric parameter construction and geometric constraints, and to provide spatiotemporal boundary conditions for radiative transfer calculation.

[0031] S2: Profile interpolation combined with elevation and observation angle to construct a scenario-based atmospheric profile parameter grid.

[0032] Based on discrete latitude and longitude grid point profiles from reanalysis atmospheric profile data, a scenario-based atmospheric parameter grid is constructed for the target scene. This reanalysis data is typically obtained from the Copernicus Climate Data Store (CDS) of the fifth-generation ECMWF atmospheric reanalysis data (ERA5). For the target location... The atmospheric profile at a given location is constructed using distance-inverse weighted interpolation to determine the local atmospheric state, and its basic form is as follows:

[0033] ;

[0034] in, This represents the atmospheric profile parameters at the source grid points. This represents the distance between the target location and the source grid point. This represents the power of distance. The original 0.25° square ERA5 atmospheric profile is interpolated using a 5×5 grid, resulting in 25 points per ERA5 profile. Interpolation points. To further improve local adaptability, based on the above spatial interpolation, terrain and observation geometry constraints are introduced: on the one hand, the surface elevation corresponding to the target grid is extracted from the digital elevation model (DEM); on the other hand, the apparent zenith angle of each target grid is corrected according to the scene center viewpoint, platform height, and the relative center position of the target point. The idea can be expressed as: the farther the target point is from the scene center, the more obvious the off-axis observation effect; the higher the target ground elevation, the shorter the effective atmospheric path.

[0035] Function: To enable each target grid used for radiative transfer solution to simultaneously possess localized atmospheric profiles, surface elevations, and observational geometric parameters, providing input conditions closer to the real scene for subsequent batch processing of the Medium Resolution Atmospheric Radiative Transfer Model (MODTRAN).

[0036] S3: Call the MODTRAN atmospheric radiative transfer model for batch calculations to obtain the radiative transfer parameters of each control node.

[0037] The scene-specific atmospheric profile parameter mesh obtained in step S2, along with aerosol types and visibility suitable for the corresponding region, is written into the MODTRAN standard input file (TAPE5). The MODTRAN atmospheric radiative transfer model is then called in batches for calculation, outputting the corresponding standard output file (TAPE7) containing hyperspectral results. This step does not directly solve for radiative transfer on a pixel-by-pixel basis for the entire image; instead, it performs batch modeling of the control mesh after spatial sampling and geometric constraints, thus significantly reducing computational complexity. In other words, this step obtains the radiative transfer response with spatially varying characteristics within the scene through "precise calculation of finite control nodes," providing a parameter basis for subsequent full-resolution image restoration.

[0038] Function: It preserves the physical rigor of the radiative transfer model while avoiding the high time cost required to run MODTRAN pixel by pixel.

[0039] S4: Perform spectral response function (SRF) convolution to extract transmittance and path radiation parameters at each band level.

[0040] The hyperspectral results output by MODTRAN are combined with the spectral response function of the sensor bands through convolution and integration to extract parameters such as atmospheric transmittance and path radiance corresponding to the infrared target bands. The convolution calculation can be expressed as:

[0041] ;

[0042] in, The spectral quantity output by MODTRAN For the first Spectral response function for each band, These are the band equivalent parameters obtained after convolution.

[0043] For transmittance, it can be obtained from the total transmittance term. Obtain the transmittance of the band The calculation formula is as follows:

[0044] ;

[0045] For path radiation, it can be determined by the path thermal radiation term. With scattering term Combining the band path radiation The calculation formula is as follows:

[0046] ;

[0047] This step converts the continuous spectral results output by the radiative transfer model into atmospheric correction parameters for the actual sensor bands. These atmospheric correction parameters are then used to generate control grid parameters based on the coverage area corresponding to the original interpolated latitude and longitude array grid points.

[0048] Function: To convert the model output spectrum into actual sensor band observation results.

[0049] S5: Surface radiance inversion is completed through adaptive fusion.

[0050] The control grid parameters obtained in step S4 are reprojected onto the original image coordinate system, and adaptive fusion interpolation is performed for each pixel of the original image to obtain an atmospheric transmittance and path radiance distribution map of the same size as the original image. Specifically, for each target pixel, two types of candidate interpolation results are constructed: one is a four-neighborhood bilinear interpolation result. Another type is the nine-neighbor distance weighted interpolation result. Then, based on the variance and gradient information of the local window of the target pixel, the fusion weights are calculated, so that flat regions tend to maintain a smooth transition, and regions with drastic changes tend to retain local structure. Its fusion form can be expressed as:

[0051] ;

[0052] in, This is for adaptive fusion weights. To construct these weights, the local variance and gradient corresponding to the four-neighbor and nine-neighbor results can be calculated separately:

[0053] ;

[0054] ;

[0055] in, The mean value within the domain. For the value of a certain point, The number of sample points in the domain. For variance For gradient, and These represent the horizontal and vertical gradients of a sample point, respectively.

[0056] The variance is compared to the gradient to form a ratio term, for example:

[0057] ;

[0058] in, It represents the variance of the four neighboring regions (a total of 4 points, up, down, left, and right). This represents the variance of the nine neighborhoods (3×3, a total of 9 points). This represents the gradient at each point within the four neighborhoods. It represents the sum (or mean) of the gradient magnitudes at each point within the nine-neighborhood. Used to characterize the change in local discreteness after the neighborhood scale is expanded. Used to characterize the change in local edge intensity after the neighborhood scale is expanded. and If the values ​​are greater than a preset threshold, the current pixel is considered to be in a region of significant local change, and the four-neighbor result is used; otherwise, the nine-neighbor result is used. Therefore, the fusion weights are... It can be represented as:

[0059] ;

[0060] in, and These are the variance ratio threshold and the gradient ratio threshold, respectively.

[0061] After mapping the ratio terms to fusion weights according to the set thresholds and completing the full-resolution reconstruction of atmospheric parameters, the observed radiance at the Earth's surface can be recovered based on the radiative transfer inversion relationship.

[0062] ;

[0063] in, The entrance pupil radiance, For process radiation, Atmospheric transmittance, This is the corrected surface radiance.

[0064] Function: To smoothly, continuously and locally adaptively propagate the high-precision radiative transfer results on the sparse control grid to every pixel of the original image, thereby achieving atmospheric correction of infrared images with both accuracy and efficiency.

[0065] This invention also provides a radiative transfer model-based adaptive atmospheric correction device for images, comprising the following modules:

[0066] 1. Data parsing module, used to read reanalysis atmospheric data, image metadata, terrain data, band response functions and reference images, and extract key parameters such as scene range, imaging time, and center view.

[0067] 2. The scene-based atmospheric parameter construction module maps discrete atmospheric profiles from reanalysis data to the target scene. It reconstructs local atmospheric profiles through spatial interpolation and incorporates terrain and observational geometric constraints by combining DEM and apparent zenith angle correction. This module successfully transforms the originally coarse-resolution, region-shared atmospheric information into control grid parameters adapted to the specific scene's changing characteristics.

[0068] 3. The radiative transfer batch solution module is used to input parameters such as atmospheric profiles, surface elevation, observation geometry, regional aerosol types, and visibility from each control grid node into TAPE5, and then call MODTRAN for batch calculations to generate TAPE7 spectral output. This module performs physical modeling and is the direct source of atmospheric correction parameters. It replaces the pixel-by-pixel calculation for the entire scene with calculations at control nodes, thereby significantly reducing the computational load.

[0069] 4. Band response convolution and parameter extraction module, used to convolve the continuous spectral results output by MODTRAN with the sensor SRF to extract the atmospheric transmittance and path radiation corresponding to each target band.

[0070] 5. Pixel-level adaptive atmospheric correction module, used to map atmospheric parameters on the control grid back to the original image resolution, and generate a smooth and edge-preserving atmospheric parameter field by pixel-by-pixel adaptive fusion of bilinear interpolation and nine-neighbor weighted interpolation, and finally complete the inversion of the observed radiance at the ground surface. Specific Implementation

[0072] like Figure 1 As shown, the present invention proposes an adaptive atmospheric correction method for infrared images based on a radiative transfer model, comprising the following steps:

[0073] S1: Input the multi-source auxiliary data corresponding to the infrared image to be corrected, and extract the atmospheric, topographic and observational geometric information of the scene to provide input conditions for the subsequent construction of scene-based atmospheric parameters.

[0074] S11 Data Input: Reads reanalysis atmospheric profile data, image metadata files, digital elevation model (DEM) data, sensor spectral response function (SRF) data, and reference projected imagery. The reanalysis data provides atmospheric profiles at discrete grid points; the image metadata provides scene extent and observation geometry; the DEM provides surface elevation information; the spectral response function is used for subsequent band convolution calculations; and the reference projected imagery is used to restore the results to the original image coordinate system.

[0075] S12 Information Extraction: Extract the coordinates of the top-left, top-right, bottom-left, and bottom-right corners of the scene, the center latitude and longitude, imaging time, and the center observation zenith angle from the image metadata, and determine the spatial coverage and center observation conditions of the target scene accordingly. This step outputs the subsequent interpolation area range, control node range, and basic parameters required for viewpoint correction.

[0076] S2: Based on the scene range and reanalysis data, construct scene-specific atmospheric parameters on the control nodes, and further modify the input conditions of each control node by combining terrain and observation geometric constraints.

[0077] S21 Profile Interpolation: Several discrete atmospheric profile nodes covering the target scene are selected from the reanalysis data, and spatial interpolation reconstruction is performed on the atmospheric profiles at the target control nodes. For the target location... Atmospheric parameters at that location were calculated using a distance-weighted method:

[0078] , ;

[0079] in, These are the atmospheric profile parameters at the source node. The distance between the target location and the source node. The distance is the power exponent. In this invention, the interpolation is not solved for a single value at once, but rather interpolated layer by layer along the entire profile, including temperature, pressure, and water vapor, to finally obtain the complete atmospheric profile corresponding to the control node.

[0080] S22 Topographic Geometric Constraints: After completing atmospheric profile interpolation, topographic and observation geometric constraints are further applied to each control node, so that different locations not only have different atmospheric profiles, but also different elevation conditions and observation zenith angle conditions, thereby making the radiative transfer input more consistent with the real scene.

[0081] S221 DEM Extraction: Based on the spatial range of the control node, the average elevation of the area is extracted from the digital elevation model and used as the surface elevation input for the control node. In this invention, instead of simply reading individual DEM pixels, the effective elevations within the corresponding range of the control node are statistically analyzed to reduce the impact of local outliers on subsequent radiative transfer solutions.

[0082] S222 Viewpoint Correction: Based on the zenith angle observed at the scene center, the satellite platform altitude, and the spatial offset of the control nodes relative to the scene center, the corrected observation zenith angle for each control node is calculated. This step ensures that different control nodes correspond to different observation path lengths, facilitating a more realistic reflection of off-axis observation effects in subsequent radiative transfer models.

[0083] S3: Write the scened atmospheric parameters on the control node into the radiative transfer model input file, and call MODTRAN to perform batch solutions to obtain the continuous spectral radiative transfer results on the control node.

[0084] S31 TAPE5 Construction: The atmospheric profiles, elevation parameters, observed zenith angles, regional aerosol types, and visibility parameters obtained in step S2 are written point by point into the TAPE5 file. Each control node corresponds to a complete set of radiative transfer model inputs, and all control nodes are organized into the input file in a batch processing manner.

[0085] S32 MODTRAN Batch Calculation: MODTRAN is invoked to solve the radiative transfer problem for each control node sequentially. Compared with directly running the radiative transfer model pixel by pixel on the entire image, this invention first performs discrete solution at the control node level, and then restores the original image resolution in subsequent steps, thereby reducing the overall computational load while maintaining the physical model constraints.

[0086] S33 TAPE7 Output: After MODTRAN calculations are completed, the TAPE7 spectral results are output. These results preserve the continuous spectral transmittance and various radiative terms corresponding to the control nodes, providing the original spectral basis for subsequent band convolution and parameter extraction.

[0087] S4: Convolve the continuous spectral results output by the radiative transfer model with the sensor spectral response function to extract the equivalent atmospheric correction parameters on the target band.

[0088] S41 Convolution Calculation: The continuous spectral result output by MODTRAN is convolved with the spectral response function of the target band to obtain the equivalent result in the corresponding band.

[0089] ;

[0090] in, The spectral quantity output by MODTRAN Let be the spectral response function of the target band. This is the equivalent result of the band obtained after convolution. In this invention, the above formula is achieved by weighted summation at discrete sampling points, that is, the spectral value and response value of each sampling wavelength point are multiplied point by point and then accumulated, and then divided by the accumulated value of the response function, thereby converting the continuous spectral result to the actual working band of the sensor.

[0091] S42 Transmittance Extraction: Convolution is performed on the total transmittance-related spectral quantities to obtain the transmittance parameters on the target band. This step compresses the continuous spectral transmittance at the control node into band transmittance that can be directly used in the correction calculation.

[0092] S43 Path Radiance Extraction: The path thermal radiation and scattered radiation related spectral quantities are combined and convolved to obtain the path radiance parameters on the target band. In this way, the radiative transfer results at the control node are uniformly converted into discrete parameters required for subsequent pixel-level atmospheric correction.

[0093] S5: Restore the band parameters on the control node to the original image resolution, and obtain the atmospheric parameter field of the same size as the original image through pixel-by-pixel adaptive fusion, and finally complete the inversion of the observed radiance at the ground surface.

[0094] S51 Adaptive Window Fusion: For each target pixel in the original image, a four-neighbor bilinear interpolation result is constructed simultaneously. Nine-neighbor weighted interpolation results And perform pixel-by-pixel fusion as follows:

[0095] ;

[0096] in, For pixel-by-pixel changing fusion weights, This refers to the final restored atmospheric parameters. In this invention, the same formula is not used uniformly across the entire scene. Instead, a weight value is calculated separately for each target pixel. If at a certain pixel... A larger value indicates that this position is more inclined to use the four-neighbor bilinear result; if A smaller value indicates that the location is more inclined to use the nine-neighbor weighted result. This allows different regions to automatically select a more suitable parameter restoration method based on local change characteristics.

[0097] S511 Bilinear Candidate Result Construction: For the control grid location of the target pixel, construct bilinear interpolation results using adjacent 2×2 control nodes. This result emphasizes the continuity of spatial transitions and is suitable for regions where parameter changes are relatively gradual.

[0098] S512 Nine Neighborhood Candidate Result Construction: Calculate the distance-weighted result for the 3×3 control nodes surrounding the target pixel. This result introduces a larger neighborhood range, which is more conducive to preserving the features of the surrounding structure when local changes are significant.

[0099] S513 fusion weight determination: Calculate separately and Variance and gradient information corresponding to the local region:

[0100] ;

[0101] ;

[0102] Then, fusion weights are generated based on the degree of local changes. :

[0103] ;

[0104] In this invention, both ratio thresholds are set to 1.35. This means that when the local variance and gradient of the nine-neighborhood are increased by more than 35% compared to the four-neighborhood, the current pixel is considered to be in a region of drastic local change, and therefore the four-neighborhood result is adopted; otherwise, the nine-neighborhood result is adopted. Experimental comparisons show that a threshold of 1.35 can better balance boundary preservation and block effect suppression.

[0105] S52 Image Reprojection: The control node parameter field is aligned with the coordinate system of the original image, and a target grid with the same size as the original image is established. This step ensures that the restored transmittance and path radiance parameters correspond one-to-one with each pixel of the original image.

[0106] S53 Inversion Calculation and Output: Using the restored transmittance and path radiance parameters, atmospheric correction is performed pixel-by-pixel on the original image to calculate the observed radiance at the Earth's surface.

[0107] ;

[0108] in, This represents the entrance pupil radiance of the original image. This represents the path radiation obtained by reconstructing the same pixel location. This represents the atmospheric transmittance recovered from the same pixel location. This represents the corrected surface radiance. This method uses the original image as a reference grid, mapping the recovered transmittance and path radiation fields pixel-by-pixel to the entire scene. Combined with the corresponding DEM elevation information, the surface radiance is inverted for each pixel based on the radiative transfer relationship, ultimately obtaining surface radiance results consistent with the spatial dimensions of the original image. For example... Figure 2 As shown, (a) presents the overall transmittance distribution, (b) presents the overall path radiance distribution, (c) presents the DEM elevation data used in parameter restoration and spatial reconstruction, and (d) presents the surface radiance image obtained by pixel-by-pixel inversion. It can be seen that both the inverted transmittance and path radiance are quite close to the actual ground cover distribution, indicating that this method can achieve continuous reconstruction of atmospheric correction parameters and rapid restoration of surface radiance information in a large-scale scene while maintaining the spatial resolution of the original image.

Claims

1. An adaptive atmospheric correction method for infrared images based on a radiative transfer model, characterized in that, Includes the following steps: S1: Acquire remote sensing images, reanalyze atmospheric profiles, digital elevation models and spectral response functions, and extract the spatial extent of the scene, imaging time and observation geometric parameters as a unified scene description; S2: Based on the discrete grid points of the reanalysis atmospheric profile, a scene-based atmospheric profile grid is constructed by inverse distance weighted interpolation, and the surface elevation of each grid point is corrected according to the digital elevation model. At the same time, the apparent zenith angle of each grid point relative to the scene center is calculated according to the observation geometric parameters to obtain the control node input parameters for solving the radiative transfer problem. S3: Write the input parameters of each control node into the standard input file of the atmospheric radiative transfer model in batches, call the atmospheric radiative transfer model to perform batch calculations, and extract the hyperspectral radiative transfer results of each control node from the standard output file. S4: Convolve the hyperspectral radiative transfer result with the spectral response function to extract the atmospheric transmittance and path radiative parameters of each control node in each infrared target band, forming a control node parameter grid. S5: Reproject the control node parameter grid onto the original image coordinate system. For each original image pixel, construct a four-neighbor bilinear interpolation result and a nine-neighbor distance-weighted interpolation result respectively. Calculate the adaptive fusion weight based on the variance and gradient information of the local window. Fuse the two interpolation results into a full-resolution atmospheric transmittance field and a path radiation field. Then, based on the radiative transfer inversion relationship between entrance pupil radiance, path radiation, and transmittance, calculate the observed radiance at the ground surface pixel by pixel.

2. The method according to claim 1, characterized in that, When constructing the scene-based atmospheric profile mesh, a 5×5 array of grid points is used to interpolate an original reanalysis atmospheric profile to obtain a complete atmospheric profile at 25 interpolation points. The complete atmospheric profile includes temperature profile, pressure profile, and water vapor profile.

3. The method according to claim 1, characterized in that, When calculating the apparent zenith angle of each grid point relative to the scene center based on the observation geometric parameters, corrections are made based on the observation zenith angle of the scene center, the satellite platform height, and the spatial offset of the grid point relative to the scene center, so that different grid points correspond to different observation path lengths.

4. The method according to claim 1, characterized in that, When extracting the atmospheric transmittance and path radiation parameters of each control node in each infrared target band, the total transmittance spectral term is convolved to obtain the band transmittance, and the sum of the path thermal radiation spectral term and the scattered radiation spectral term is convolved to obtain the band path radiation.

5. The method according to claim 1, characterized in that, When calculating the adaptive fusion weight, the local variance and local gradient magnitude corresponding to the four-neighbor result and the nine-neighbor result are calculated respectively. Then, the variance ratio and gradient ratio of the nine-neighbor result and the four-neighbor result are calculated. When the variance ratio and gradient ratio are both greater than the preset threshold, the fusion weight is set to 1 to use the four-neighbor bilinear interpolation result; otherwise, the fusion weight is set to 0 to use the nine-neighbor distance-weighted interpolation result.

6. The method according to claim 5, characterized in that, The preset threshold is 1.35, which means that when the local variance and local gradient magnitude of the nine-neighborhood relative to the four-neighborhood both increase by more than 35%, the current pixel is determined to be located in a region of drastic local change.

7. The method according to claim 1, characterized in that, The formula used for calculating the observed radiance at the Earth's surface pixel by pixel is: ； in, The entrance pupil radiance, For process radiation, Atmospheric transmittance, This is the corrected surface radiance.

8. An adaptive atmospheric correction device for infrared images based on a radiative transfer model, characterized in that, include: The parameter acquisition module acquires remote sensing images, reanalysis atmospheric profiles, digital elevation models and spectral response functions, and extracts the spatial extent of the scene, imaging time and observation geometric parameters as a unified scene description. The parameter solving module constructs a scene-based atmospheric profile grid based on the discrete grid points of the reanalysis atmospheric profile through inverse distance weighted interpolation, corrects the surface elevation of each grid point according to the digital elevation model, and calculates the apparent zenith angle of each grid point relative to the scene center according to the observation geometric parameters to obtain the control node input parameters for solving radiative transfer. The result extraction module writes the input parameters of each control node into the standard input file of the atmospheric radiative transfer model in batches, calls the atmospheric radiative transfer model to perform batch calculations, and extracts the hyperspectral radiative transfer results of each control node from the standard output file. The parameter grid extraction module convolves the hyperspectral radiative transfer result with the spectral response function to extract the atmospheric transmittance and path radiative parameters of each control node in each infrared target band, forming a control node parameter grid. The radiance calculation module reprojects the control node parameter grid onto the original image coordinate system. For each original image pixel, it constructs a four-neighbor bilinear interpolation result and a nine-neighbor distance-weighted interpolation result, respectively. Based on the variance and gradient information of the local window, it calculates an adaptive fusion weight and fuses the two interpolation results into a full-resolution atmospheric transmittance field and a path radiation field. Then, based on the radiative transfer inversion relationship between entrance pupil radiance, path radiation, and transmittance, it calculates the observed radiance at the ground surface pixel by pixel.

9. An electronic device, characterized in that, include: One or more processors; A memory for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the method of any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, It stores executable instructions that, when executed by a processor, cause the processor to perform the method described in any one of claims 1 to 7.

Images