Modulation transfer function controlled remote sensing image radiation high-fidelity adaptive degradation method

Abstract

This invention discloses a high-fidelity adaptive degradation method for radiometric data in remote sensing images controlled by modulation transfer function. First, channel-independent radiometric fidelity normalization is performed for each band. Then, candidate fuzzy kernel parameters are generated through a multi-kernel family sampling mechanism and content-adaptive fuzzy parameter mapping. Next, the parameters are physically constrained and truncated based on the spatial response information or modulation transfer function information disclosed by the target sensor. Subsequently, fuzzy convolution, downsampling, and optional noise superposition are performed. Finally, the original radiometric scale of each band is restored through inverse normalization to obtain a low-resolution degraded image. This invention can better simulate the spatial fuzziness characteristics and frequency attenuation laws in real remote sensing sensor imaging, exhibiting good stability and practicality. It is suitable for multispectral remote sensing image degradation modeling, training sample construction, and related remote sensing image processing applications, providing a more reliable data foundation for remote sensing image super-resolution reconstruction, image enhancement, and related downstream tasks.

Description

Technical Field

[0001] This invention relates to a remote sensing image degradation method, specifically a high-fidelity adaptive degradation method for remote sensing images that combines multi-kernel family sampling, content-adaptive fuzzy mapping, and modulation transfer function control constrained by sensor imaging characteristics, belonging to the field of remote sensing image processing and image reconstruction technology. Background Technology

[0002] Super-resolution reconstruction of multispectral remote sensing images aims to recover details from high-resolution images from low-resolution remote sensing images, which is of great significance for improving the accuracy of remote sensing image interpretation, ground feature identification capabilities, and resource and environmental monitoring. In supervised learning-based super-resolution reconstruction tasks, it is usually necessary to construct paired training samples between low-resolution images and their corresponding high-resolution images. Therefore, how to reasonably degrade high-resolution multispectral remote sensing images to generate low-resolution samples that more closely resemble the actual sensor observation process is a crucial factor affecting the training effect of super-resolution models.

[0003] In existing technologies, the degradation methods used to construct training samples are usually quite simple, often employing bicubic downsampling, fixed Gaussian blur, or a combination of both to generate low-resolution images. While these methods are easy to implement, their degradation process differs significantly from the imaging mechanism of real remote sensing sensors. This results in the generated low-resolution samples failing to maintain consistency with real observations in terms of spatial blur morphology, frequency domain attenuation characteristics, and radiometric response, thus limiting the training effectiveness and generalization ability of super-resolution reconstruction models.

[0004] For multispectral remote sensing images, different bands typically have different radiometric calibration ranges, grayscale distributions, and dynamic ranges. Most existing degradation methods follow a common uniform normalization strategy found in natural image processing, processing all channels holistically. This approach easily overlooks the radiometric differences between different bands, causing a shift in the band response relationship before and after degradation. This leads to cross-band radiometric distortion and artifacts, hindering the accurate learning of multi-band information by subsequent reconstruction models.

[0005] The sources of blur degradation in real remote sensing imaging are complex, potentially originating from the point spread response of the optical system, or related to factors such as platform motion, attitude jitter, and differences in the imaging link, exhibiting various blur patterns including directional, tail attenuation variations, and flat-top responses. However, existing methods typically employ only a single Gaussian kernel to approximate the degradation process, which is insufficient to fully cover the aforementioned complex blur patterns, thus resulting in inadequate realism in the generated samples.

[0006] Furthermore, existing degradation methods typically employ spatially uniform blur parameters, failing to consider the differences in texture complexity and structural sensitivity among different land cover regions. In reality, textured regions, regions with prominent edges, and flat regions exhibit varying degrees of sensitivity to blur degradation. If a fixed blur intensity is uniformly applied, it is difficult to effectively simulate the characteristics of spatially non-uniform degradation in real remote sensing images.

[0007] In addition, the spatial blurring characteristics of real satellite sensors are usually constrained by the modulation transfer function (MTF) or spatial response characteristics of their imaging system. However, most existing degradation modeling methods lack a correspondence between the blur kernel parameters and the physical imaging characteristics of the target sensor, resulting in significant differences between the generated low-resolution samples and real observations in terms of high-frequency attenuation trends, edge loss, and texture preservation characteristics.

[0008] In summary, existing remote sensing image degradation modeling methods still have the following shortcomings: first, they fail to adequately maintain the radiometric consistency of each band in multispectral images; second, they struggle to characterize multiple types of blur degradation in remote sensing imaging; third, they lack adaptive blur modeling based on image content differences; and fourth, they lack physical constraints related to the imaging characteristics of the target sensor. Therefore, how to simultaneously consider multi-band radiometric fidelity, the diversity of blur morphologies, adaptive degradation characteristics based on content, and the constraints of real sensor spatial response to construct a degradation model that better reflects the actual remote sensing imaging process, thereby generating more realistic low-resolution training samples, has become an urgent problem to be solved in this field. Summary of the Invention

[0009] To address the problems existing in the prior art, this invention provides a high-fidelity adaptive degradation method for radiometric data of remote sensing images controlled by modulation transfer function. This method can improve the consistency between the degraded samples and real sensor observations in terms of spatial blurring morphology and frequency domain attenuation characteristics while preserving the radiometric characteristics of multispectral remote sensing images. This provides more realistic training samples for the super-resolution reconstruction model of remote sensing images and enhances the reconstruction performance and generalization ability of the model.

[0010] To achieve the above objectives, the modulation transfer function-controlled remote sensing image radiometric high-fidelity adaptive degradation method specifically includes the following steps:

[0011] Step 1: Obtain high-resolution multispectral remote sensing images from the target area or target dataset as input. The high-resolution multispectral remote sensing images are represented as follows: ,in: Represents the real number field. This indicates the input high-resolution multispectral remote sensing image. Indicates the number of spectral channels. Indicates the height of the input image. Indicates the width of the input image;

[0012] Step 2: Perform channel normalization processing on each band of the high-resolution multispectral remote sensing image to obtain the normalized image.

[0013] Step 3: Determine the blur kernel type based on the multi-kernel family sampling mechanism, and perform content-adaptive blur parameter mapping based on image content features to obtain blur kernel parameters;

[0014] Step 4: Constrain and trim the fuzzy kernel parameters according to the imaging characteristics of the target sensor. First, determine the range of the equivalent point spread function parameters corresponding to the target sensor based on the publicly available spatial response information or modulation transfer function information of the target sensor. Then, restrict the parameters obtained by content adaptive fuzzy parameter mapping to this physical range.

[0015] Step 5: Perform blur convolution degradation on the normalized image using the cropped blur kernel;

[0016] Step 6: Perform downsampling on the image after blurring and convolution degradation, with a scaling factor of [value missing]. Downsampling yields low-resolution images;

[0017] Step 7: Selectively add noise terms to the downsampled low-resolution image;

[0018] Step 8, the normalized domain image after downsampling and optional noise overlay is denoted as ,right Perform channel-independent radiometric inverse normalization and constrain the inverse normalization results within the radiometric range of the corresponding band to obtain low-resolution degraded images. ,in: This indicates a low-resolution degraded image output. and These represent the height and width of the output image, respectively. , .

[0019] Furthermore, the specific process of Step 2 is as follows:

[0020] For the first input image For each band, the minimum radiation value is first calculated in the spatial dimension. and maximum radiation value , means as follows:

[0021]

[0022] In the formula: Indicates the input image number The spatial location of each band Pixel value at; Indicates the first Minimum radiation value for each band; Indicates the first Maximum radiation value of each band;

[0023] Then for the first Each band undergoes independent channel normalization processing, as shown below:

[0024]

[0025] In the formula, Indicates the image after normalization. Each band in position Pixel value at; This is a preset minimum positive number to prevent the denominator from being zero; This indicates an operation to select the larger value.

[0026] Furthermore, the normalization process in Step 2 also includes outlier robustness processing, which employs a robust mapping method based on quantiles, replacing the minimum radiation value with the preset values ​​corresponding to low and high quantiles, respectively. and maximum radiation value Normalization and denormalization are performed.

[0027] Furthermore, the specific process of Step 3 is as follows:

[0028] Step 3-1, Construct a multi-core family sampling mechanism:

[0029] Preset fuzzy kernel family Recorded as:

[0030]

[0031] In the formula: Indicates candidate fuzzy kernel type I; Indicates candidate fuzzy kernel type II; Indicates candidate fuzzy kernel type III;

[0032] The fuzzy kernel type is randomly sampled according to a mixed distribution, as follows:

[0033]

[0034] In the formula: , , They represent , , The sampling weights, and satisfying ; Indicates the type of fuzzy kernel obtained from sampling; This represents the specific fuzzy kernel instantiated based on the fuzzy kernel type;

[0035] Step 3-2, Adaptive fuzzy parameter mapping based on image content features: Adopting a content-adaptive fuzzy parameter mapping strategy, the sampling range of fuzzy parameters is dynamically adjusted according to the texture complexity of the input image.

[0036] Furthermore, the preset fuzzy kernel family in Step 3-1 middle, Anisotropic Gaussian kernel, For the generalized radial exponent kernel, For a flat-topped radial core, the mathematical expression is as follows:

[0037]

[0038]

[0039]

[0040] In the formula: Indicates the coordinate position of the fuzzy kernel. ; Indicates the rotation angle A defined two-dimensional rotation matrix; and These represent the diffusion scale parameters along the main and secondary directions, respectively. is the normalization constant for the anisotropic Gaussian kernel; The Euclidean norm of a vector; Indicates the radial scale parameter; Indicates the tail shape parameters; , , These are the normalization constants for the anisotropic Gaussian kernel, the radial exponential kernel, and the flat-top radial kernel, respectively.

[0041] Furthermore, the specific process of Step 3-2 is as follows:

[0042] Step 3-2-1, input image Convert to grayscale reference image Grayscale reference image By using the preset channel set The pixel mean value of multiple channels in the image is obtained as follows:

[0043]

[0044] In the formula: Represents a grayscale reference image; Indicates a preset set of channels; Represents the preset channel set The number of channels in; Indicates the first input image One channel. A preferred embodiment uses a preset channel set. It is a set of visible light channels;

[0045] Step 3-2-2: Calculate the grayscale reference images using the Sobel operator. Horizontal and vertical gradient components and And calculate the gradient magnitude of the entire image, as shown below:

[0046]

[0047] In the formula: Represents grayscale reference image In spatial location Gradient magnitude at; Represents grayscale reference image In spatial location The gradient component in the horizontal direction; Represents grayscale reference image In spatial location The gradient component in the vertical direction;

[0048] Step 3-2-3: Take the average gradient magnitude of the entire image and use it as the texture score of the current sample. , means as follows:

[0049]

[0050] In the formula, This represents the texture score of the current input sample;

[0051] Step 3-2-4: Normalize the texture scores Mapped to a blur parameter sampling interval to control the intensity range of the blur kernel, texture score The sampling interval of the fuzzy parameter satisfies the following monotonically decreasing relationship:

[0052]

[0053] In the formula: This represents the sampling interval of the fuzzy parameters corresponding to the current sample; This represents a mapping function from texture scores to the range of blur parameters. This indicates that the mapping function is monotonically decreasing with respect to the texture score.

[0054] Furthermore, the specific process of Step 4 is as follows:

[0055] Step 4-1: Construct parameter constraints based on the imaging characteristics of the target sensor:

[0056] The equivalent point spread function of the target sensor is approximated by a two-dimensional anisotropic Gaussian function, as follows:

[0057]

[0058] In the formula: This represents the point spread function of the target sensor in the spatial domain. and Represents spatial coordinates; and These represent the diffusion parameters along the horizontal and vertical directions, respectively.

[0059] Its corresponding frequency domain modulation transfer function (MTF) is approximately expressed as follows:

[0060]

[0061] In the formula: Represents the frequency domain modulation transfer function; and These represent the spatial frequency coordinates in the horizontal and vertical directions, respectively.

[0062] The frequency domain modulation transfer function (MTF) is represented in a one-dimensional equivalent form as follows:

[0063]

[0064] In the formula: Represents frequency in one-dimensional space; Represents the one-dimensional equivalent diffusion parameter;

[0065] Based on the MTF range of the target sensor at the Nyquist frequency, the one-dimensional equivalent diffusion parameters are obtained by inversion. The physically feasible range;

[0066] Step 4-2, Parameter Trimming and Physical Consistency Constraints:

[0067] When the target sensor mode is enabled, the parameters obtained by content-adaptive fuzzy mapping are intersected with the physical constraint range of the target sensor and then clipped, including at least:

[0068]

[0069]

[0070] In the formula: and These represent the horizontal and vertical diffusion parameters obtained from the content-adaptive fuzzy mapping, respectively. , , , These represent the upper and lower bounds of the physical constraint range of the target sensor, respectively. This represents a clipping function used to restrict input parameters to a given range.

[0071] Furthermore, in Step 8, regarding... Perform independent radial denormalization of the execution channel, as follows:

[0072]

[0073] In the formula: This represents the output pixel value after denormalization; Indicates the degenerate th term within the normalization domain. Each band in position Pixel value at; This indicates that the input value will be clipped to a range. Inside; and They represent the first The maximum and minimum radiation values ​​of the original radiation range of each band.

[0074] Furthermore, in Step 6, bicubic interpolation downsampling is used.

[0075] Furthermore, in Step 7, the superimposed noise term is at least one of Gaussian noise, Poisson noise, or a combination of noise corresponding to the noise characteristics of the target sensor.

[0076] Compared with existing technologies, the modulation transfer function-controlled remote sensing image radiometric high-fidelity adaptive degradation method has the following advantages:

[0077] 1. This invention processes each band of a multispectral remote sensing image separately through independent normalization and denormalization mechanisms for each band of the image. This effectively maintains the consistency of the original radiation range of each band and reduces cross-band radiation shift and artifact problems caused by uniform normalization.

[0078] 2. This invention constructs a multi-kernel family including anisotropic Gaussian kernels, generalized radial exponential kernels, and flat-top radial kernels, and combines it with a hybrid distribution sampling mechanism, which can more fully cover various blur patterns in remote sensing imaging and improve the ability of the degradation model to represent complex blur patterns.

[0079] 3. This invention uses a content-adaptive fuzzy parameter mapping mechanism to dynamically adjust the fuzz intensity based on the texture complexity of the input image, so that texture-rich areas and texture-flat areas correspond to different degrees of degradation, thus better conforming to the characteristics of non-uniform spatial degradation of real remote sensing images.

[0080] 4. This invention introduces a target sensor modulation transfer function constraint mechanism and uses publicly available spatial response information or modulation transfer function information to physically trim the fuzzy kernel parameters, making the synthesized degraded samples closer to the real sensor observation results in terms of frequency domain attenuation trend, edge loss degree and texture preservation characteristics.

[0081] 5. This invention organically combines spectral radiometric fidelity, multi-kernel family fuzzy modeling, content-adaptive parameter mapping, and sensor imaging characteristic constraints to generate more realistic low-resolution degraded samples, thereby providing a more reliable data foundation for remote sensing image super-resolution reconstruction, image enhancement, and related downstream tasks. Attached Figure Description

[0082] Figure 1 This is a schematic diagram of the overall process of the present invention;

[0083] Figure 2 This is a schematic diagram comparing the degradation results of the method of this invention with those of traditional degradation methods on the RGB band of a public dataset;

[0084] Figure 3 This diagram illustrates a comparison of degradation results using the method of this invention and traditional degradation methods on a multi-band dataset of a custom Gaofen-1 satellite imagery dataset. Detailed Implementation

[0085] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0086] This modulation transfer function-controlled high-fidelity adaptive degradation method for remote sensing images aims to address the problems in existing multispectral remote sensing image degradation modeling, such as distorted radiative ranges across different bands, insufficient ability to represent fuzzy types, difficulty in adaptively adjusting degradation intensity based on image content, and inconsistencies between degradation results and actual sensor imaging characteristics. First, channel-independent radiometric fidelity normalization is performed for each band. Then, candidate fuzzy kernel parameters are generated through a multi-kernel family sampling mechanism and content-adaptive fuzzy parameter mapping. Next, the parameters are physically constrained and truncated based on the spatial response information or modulation transfer function information publicly available from the target sensor. Subsequently, fuzzy convolution, downsampling, and optional noise superposition are performed. Finally, inverse normalization is used to restore the original radiative scale of each band, resulting in a low-resolution degraded image.

[0087] like Figure 1 As shown, the modulation transfer function-controlled remote sensing image radiometric high-fidelity adaptive degradation method specifically includes the following steps:

[0088] Step 1: Obtain high-resolution multispectral remote sensing images from the target area or target dataset as input. The high-resolution multispectral remote sensing images are represented as follows: ,in: Represents the real number field. This indicates the input high-resolution multispectral remote sensing image. Indicates the number of spectral channels. Indicates the height of the input image. Indicates the width of the input image.

[0089] High-resolution multispectral remote sensing imagery of the target area or dataset is acquired as input. This imagery can originate from publicly available remote sensing datasets, self-built satellite remote sensing datasets, aerial remote sensing imagery, or other Earth observation data with multispectral band information. The input imagery typically contains multiple spectral channels, such as blue, green, red, and near-infrared bands. Different bands often correspond to different radiometric calibration ranges and response characteristics. This invention does not limit the specific spatial resolution, number of spectral channels, or sensor source of the input imagery, as long as it can provide multi-band remote sensing observation data.

[0090] Step 2: Perform channel normalization processing on each band of the high-resolution multispectral remote sensing image to obtain the normalized image.

[0091] Since different bands of multispectral remote sensing images typically have different radiometric calibration ranges, grayscale distributions, and dynamic ranges, directly applying a uniform normalization method can easily cause interference between radiometric information from different bands, leading to cross-band shifts and radiometric distortion. Therefore, this invention performs channel normalization processing separately for each band of the image, as follows:

[0092] For the first input image For each band, the minimum radiation value is first calculated in the spatial dimension. and maximum radiation value , means as follows:

[0093]

[0094] In the formula: Indicates the input image number The spatial location of each band Pixel value at; Indicates the first Minimum radiation value for each band; Indicates the first The maximum radiation value of each band.

[0095] To obtain the minimum radiation value and maximum radiation value Afterwards, regarding the first Each band undergoes independent channel normalization processing, as shown below:

[0096]

[0097] In the formula, Indicates the image after normalization. Each band in position Pixel value at; This is a preset minimum positive number to prevent the denominator from being zero; This indicates an operation to select the larger value.

[0098] Through the above processing, each band completes numerical normalization within its own radiation dynamic range, thereby effectively avoiding the destruction of multi-band radiation relationships by the unified normalization strategy.

[0099] To improve the robustness of the normalization process to outliers, the normalization process can also include outlier robustness processing. Outlier robustness processing can adopt a robust mapping method based on quantiles, replacing the minimum radiation value with the values ​​corresponding to preset low quantiles and high quantiles, respectively. and maximum radiation value Normalization and denormalization are performed to reduce the impact of isolated outliers on radiation range estimation.

[0100] Step 3: Determine the blur kernel type based on the multi-kernel family sampling mechanism, and perform content adaptive blur parameter mapping based on image content features to obtain blur kernel parameters.

[0101] To more fully simulate the complex and diverse sources of blur during remote sensing imaging, this invention employs a multi-kernel sampling mechanism, randomly selecting a degradation kernel type from multiple blur kernel types with complementary representation capabilities. Simultaneously, based on the content characteristics of the input image, the range of blur parameters is adaptively adjusted to match the degradation intensity with the complexity of the scene texture. Specifically:

[0102] Step 3-1, Construct a multi-core family sampling mechanism:

[0103] Preset fuzzy kernel family Recorded as:

[0104]

[0105] In the formula: Indicates candidate fuzzy kernel type I, in the example Anisotropic Gaussian kernels are used to characterize directional blur in remote sensing imaging; Indicates candidate fuzzy kernel type II, in the example It is a generalized radial exponential kernel used to characterize fuzziness with variable tail decay morphology; Indicates candidate fuzzy kernel type III, in the example It is a flat-topped radial core, used to characterize the fuzziness of a relatively flat diffusion response.

[0106] The fuzzy kernel type is randomly sampled according to a mixed distribution, as follows:

[0107]

[0108] In the formula: , , They represent , , The sampling weights, and satisfying ; Indicates the type of fuzzy kernel obtained from sampling; This indicates the specific fuzzy kernel instantiated based on the fuzzy kernel type.

[0109] Anisotropic Gaussian kernel The mathematical expression is as follows:

[0110]

[0111] In the formula: Indicates the coordinate position of the fuzzy kernel. ; Indicates the rotation angle A defined two-dimensional rotation matrix; and These represent the diffusion scale parameters along the main and secondary directions, respectively. is the normalization constant of the anisotropic Gaussian kernel, used to ensure that the sum of the fuzzy kernel coefficients is 1.

[0112] Generalized radial exponent kernel Flat-top radial core The mathematical expressions for them are as follows:

[0113]

[0114]

[0115] In the formula: Indicates the coordinate position of the fuzzy kernel. ; The Euclidean norm of a vector; Indicates the radial scale parameter; Indicates the tail shape parameters; and These are the normalization constants for the radial exponential kernel and the flat-top radial kernel, respectively.

[0116] The above multi-core family design can cover a variety of blur response modes commonly found in remote sensing imaging with low parameter complexity.

[0117] Step 3-2, Adaptive mapping of fuzzy parameters based on image content features:

[0118] Considering the significant differences in texture complexity among different remote sensing scenarios, applying the same blur parameter range to all input samples would fail to reflect the varying sensitivities of different regions to blur degradation in real-world imaging. Therefore, this invention further employs a content-adaptive blur parameter mapping strategy, dynamically adjusting the blur parameter sampling interval based on the texture complexity of the input image. Specifically:

[0119] Step 3-2-1: First, input the image. Convert to grayscale reference image Grayscale reference image By using the preset channel set The pixel mean value of multiple channels in the image is obtained as follows:

[0120]

[0121] In the formula: Represents a grayscale reference image; Indicates a preset set of channels; Represents the preset channel set The number of channels in; Indicates the first input image One channel. A preferred embodiment uses a preset channel set. It is a set of visible light channels.

[0122] Step 3-2-2, after obtaining the grayscale reference image Then, the Sobel operator is used to calculate its gradient components in the horizontal and vertical directions, respectively. and Furthermore, the gradient magnitude of the entire image is calculated, as shown below:

[0123]

[0124] In the formula: Represents grayscale reference image In spatial location Gradient magnitude at; Represents grayscale reference image In spatial location The gradient component in the horizontal direction; Represents grayscale reference image In spatial location The gradient component in the vertical direction.

[0125] Step 3-2-3: Take the average gradient magnitude of the entire image and use it as the texture score of the current sample. , means as follows:

[0126]

[0127] In the formula, This represents the texture score of the current input sample, used to characterize the overall texture complexity of the current input image sample.

[0128] Step 3-2-4: Normalize the texture scores Mapped to a blur parameter sampling interval to control the intensity range of the blur kernel, texture score. The sampling interval of the fuzzy parameter satisfies the following monotonically decreasing relationship:

[0129]

[0130] In the formula: This represents the sampling interval of the fuzzy parameters corresponding to the current sample; This represents a mapping function from texture scores to the range of blur parameters. This indicates that the mapping function is monotonically decreasing with respect to the texture score.

[0131] This mapping relationship indicates that when the texture score of the input image... When the value is high, the blur parameter range is smaller, corresponding to a lighter blur degradation, in order to preserve the detail information of texture-rich areas; when the texture score is high... When the value is low, the range of blur parameter values ​​is large, corresponding to stronger blur degradation, which is more consistent with the response characteristics of smooth regions in actual imaging.

[0132] This content-adaptive mapping process enables the blur intensity to be dynamically adjusted as the image content changes, avoiding the degradation and distortion problems caused by uniform blur parameters. This allows the generated degradation samples to better conform to the spatial non-uniform degradation patterns in real remote sensing images.

[0133] Step 4: Constrain and trim the blur kernel parameters according to the imaging characteristics of the target sensor.

[0134] While content-based fuzzy parameter mapping can improve the adaptability of the degradation process to scene changes, without further physical constraints related to the real sensor, the fuzzy parameters, although reflecting scene differences, may still deviate from the imaging response range of the real sensor. Therefore, this invention further introduces a constraint mechanism based on the imaging characteristics of the target sensor to further prune the fuzzy parameters. Based on the publicly available spatial response information or modulation transfer function information of the target sensor, the equivalent point spread function parameter range corresponding to the target sensor is determined. The equivalent point spread function parameter range characterizes the allowable spatial spread of the target sensor, making the degraded image more closely resemble the observation results of the real sensor in terms of frequency domain attenuation characteristics. The parameters obtained from the content-adaptive fuzzy parameter mapping are then limited to this physical range. Specifically:

[0135] Step 4-1: Construct parameter constraints based on the imaging characteristics of the target sensor:

[0136] The embodiment uses a two-dimensional anisotropic Gaussian function to approximate the equivalent point spread function of the target sensor, as shown below:

[0137]

[0138] In the formula: This represents the point spread function of the target sensor in the spatial domain. and Represents spatial coordinates; and These represent the diffusion parameters along the horizontal and vertical directions, respectively.

[0139] Its corresponding frequency domain modulation transfer function (MTF) is approximately expressed as follows:

[0140]

[0141] In the formula: Represents the frequency domain modulation transfer function; and These represent the spatial frequency coordinates in the horizontal and vertical directions, respectively.

[0142] To facilitate the estimation of parameter range using publicly available sensor specifications, the frequency domain modulation transfer function (MTF) can be represented in a one-dimensional equivalent form as follows:

[0143]

[0144] In the formula: Represents frequency in one-dimensional space; This represents the one-dimensional equivalent diffusion parameter.

[0145] Based on the MTF range of the target sensor at the Nyquist frequency, the one-dimensional equivalent diffusion parameters can be further obtained through inversion. The physical feasible range is determined by converting publicly available sensor imaging metrics into constraints on the degraded kernel parameters.

[0146] Step 4-2, Parameter Trimming and Physical Consistency Constraints:

[0147] When the target sensor mode is enabled, the parameters obtained by content-adaptive fuzzy mapping are intersected with the physical constraint range of the target sensor and then clipped, including at least:

[0148]

[0149]

[0150] In the formula: and These represent the horizontal and vertical diffusion parameters obtained from the content-adaptive fuzzy mapping, respectively. , , , These represent the upper and lower bounds of the physical constraint range of the target sensor, respectively. This represents a clipping function used to restrict input parameters to a given range.

[0151] By using the above cropping, the blur kernel parameters can simultaneously satisfy the adaptive characteristics of image content and the imaging constraints of the target sensor.

[0152] Step 5: Perform blur convolution degradation on the normalized image using the cropped blur kernel to simulate the blurring effect caused by factors such as spatial diffusion of the optical system, platform motion, and imaging link during remote sensing imaging.

[0153] Step 6: Perform downsampling on the image after blurring and convolution degradation, with a scaling factor of [value missing]. The image is downsampled to obtain a low-resolution image.

[0154] The preferred downsampling method in this embodiment is bicubic interpolation downsampling.

[0155] Step 7: Selectively add noise terms to the downsampled low-resolution image.

[0156] The noise term can be at least one of Gaussian noise, Poisson noise, or a combination of noise corresponding to the noise characteristics of the target sensor.

[0157] Step 8, the normalized domain image after downsampling and optional noise overlay is denoted as ,right Perform channel-independent radiometric inverse normalization and constrain the inverse normalization results within the radiometric range of the corresponding band to obtain low-resolution degraded images. ,in: This indicates a low-resolution degraded image output. and These represent the height and width of the output image, respectively. , , This represents the downsampling scale factor.

[0158] right Perform independent radial denormalization of the execution channel, as follows:

[0159]

[0160] In the formula: This represents the output pixel value after denormalization; Indicates the degenerate th term within the normalization domain. Each band in position Pixel value at; This indicates that the input value will be clipped to a range. Inside; and They represent the first The maximum and minimum radiation values ​​of the original radiation range of each band.

[0161] To verify the effectiveness of the modulation transfer function-controlled high-fidelity adaptive degradation method for remote sensing images in generating degradation samples, this invention's method was compared with traditional degradation methods and existing image degradation methods, including LR, Bicubic, BSRGAN, Real-ESRGAN, and SEN2NAIP, to generate multispectral remote sensing image degradation samples on common RGB band scenes in publicly available datasets. The comparison results are as follows: Figure 2 As shown.

[0162] Depend on Figure 2 It can be seen that the LR method only performs direct low-resolution scaling on the original image. Although it can reduce spatial resolution, it lacks the characterization of optical system blurring, sensor frequency response attenuation, and ground feature edge diffusion phenomena in real remote sensing imaging. Therefore, the generated results are generally idealized, with harsh edge transitions and insufficient texture attenuation. The Bicubic method achieves scale transformation through bicubic interpolation, which has a certain smoothing effect. However, its degradation process is mainly determined by a fixed interpolation function and cannot adaptively adjust the degradation intensity according to different ground feature structures and spatial frequency components in the remote sensing image. This results in residual or overly smoothed phenomena that do not conform to the real imaging rules for high-frequency structures such as building edges, road boundaries, and runway lines.

[0163] While natural image degradation methods such as BSRGAN and Real-ESRGAN introduce complex forms of degradation including blurring, noise, and compression, their degradation models are primarily designed for ordinary natural image distributions and do not fully consider the modulation transfer function characteristics in the imaging chain of remote sensing sensors. In remote sensing scenarios, these methods are prone to introducing unstable pseudo-textures, color noise, or localized graininess. For example, in... Figure 2 In the athletic fields, building complexes, farmland, and river areas, the Real-ESRGAN generation results show obvious random noise and non-realistic texture enhancement, causing deviations in local spatial structure from the real low-resolution remote sensing image. Although the SEN2NAIP method is designed for remote sensing image scenes and outperforms ordinary natural image degradation methods in terms of overall color and spatial degradation performance, its degradation process still depends more on specific data distributions and is difficult to precisely control for the frequency domain response characteristics of different remote sensing images.

[0164] In contrast, the method of this invention uses the modulation transfer function as a constraint to model the spatial frequency attenuation during the degradation process of multispectral remote sensing images, enabling the generated low-resolution degraded images to more closely resemble the actual remote sensing observations in terms of visual appearance and imaging mechanism. Specifically, in densely built-up areas, the images generated by the method of this invention can better represent the natural diffusion characteristics of roof edges, road outlines, and boundaries between features, avoiding the single smoothing of the Bicubic method and the noise artifacts of the Real-ESRGAN method; in areas with regular structures such as athletic fields, the method of this invention can maintain the overall outline that should exist in low-resolution images, while reasonably weakening high-frequency details such as track lines and grass areas, so that the degradation results conform to the texture attenuation law after the sensor imaging resolution is reduced; in natural feature areas such as farmland, woodland, and rivers, the method of this invention can maintain the continuity of regional tone and radiation distribution, making edge blurring, texture fading, and feature transitions more natural.

[0165] Using the method of this invention, along with LR, Bicubic, BSRGAN, Real-ESRGAN, and SEN2NAIP methods, multispectral remote sensing image degradation samples were generated under multi-band conditions of a custom Gaofen-1 satellite multi-band image dataset. The comparison results are as follows: Figure 3 As shown.

[0166] Depend on Figure 3It can be seen that in real Gaofen-1 satellite multi-band imagery scenes, different land cover types have more complex spectral responses and spatial structure characteristics. For example, farmland, rivers, roads, buildings, and shaded areas have different radiation distributions and edge representations in different bands. Therefore, degradation sample generation methods not only need to simulate the blurring effect caused by reduced spatial resolution, but also need to maintain the consistency of radiation relationships and degradation processes across multiple bands. Although the LR and Bicubic methods can achieve basic scale changes, their degradation mechanisms are relatively simple and difficult to reflect the response attenuation characteristics of real remote sensing sensors at different spatial frequencies. Figure 3 In the results, Bicubic generally exhibits a strong smoothing effect, but the diffusion patterns of some roads, field boundaries, and water edges are not natural enough, making it difficult to accurately simulate the imaging degradation process in real low-resolution multispectral images.

[0167] BSRGAN and Real-ESRGAN are relatively less adaptable to multi-band remote sensing imagery. In particular, the Real-ESRGAN method generates significant granular noise and irregular pseudo-textures in some areas, leading to anomalous perturbations in the local radiative distribution of the image. This phenomenon is especially detrimental to multispectral remote sensing imagery because multi-band data depends not only on the visual quality of individual bands but also on stable radiometric correspondences between different bands. If random noise or inconsistent texture perturbations between bands are introduced during degradation, it may disrupt the spectral correlation between different bands, thereby affecting the reliability of subsequent tasks such as land cover classification, target identification, index calculation, and super-resolution reconstruction. The SEN2NAIP method has certain advantages in remote sensing data degradation, but from... Figure 3 The results show that in some high-frequency structural regions, there are still problems such as insufficient edge preservation or relatively fixed local blur patterns, making it difficult to achieve sufficient adaptive degradation for different scene content.

[0168] This invention introduces a modulation transfer function control mechanism to explicitly integrate the response characteristics of remote sensing imaging systems to different spatial frequency components into the degradation sample generation process, making the degradation results more physically plausible and scene-adaptable under multi-band conditions. For areas with obvious directionality and edge structure, such as farmland grids, road boundaries, and water shorelines, this invention can maintain reasonable structural contours while reducing spatial resolution, making edge transitions present a gradual diffusion state as in real imaging, without producing over-sharpening, random noise, or unnatural block artifacts. For large areas of water, vegetation, and dark features, this invention can maintain the smoothness of the radiation distribution within the area and the relative relationships between bands, avoiding local brightness anomalies, color shifts, or spectral response distortions caused by degradation processing.

[0169] Furthermore, the method of this invention has good multi-scenario adaptability. For Figure 3 In regions with varying spatial texture complexity, the method of this invention can adaptively adjust degradation performance based on image content and frequency distribution: in high-frequency texture-rich regions, it can reasonably suppress details exceeding the resolution capability of low-resolution sensors; in low-frequency homogeneous regions, it can maintain stable radiometric continuity and spatial smoothness. This characteristic ensures that the generated samples are neither overly idealized nor introduce random artifacts unrelated to real remote sensing imaging, thus better conforming to the degradation patterns observed in the actual imaging process of the Gaofen-1 multispectral sensor.

[0170] The above implementation results demonstrate that the method of this invention can generate low-resolution degraded samples with good realism and consistency under different scene types and different band conditions. For textured regions, the method of this invention can avoid abnormal loss of details caused by excessive blurring; for smooth regions, it can suppress unreasonable high-frequency residues or insufficient degradation, thereby making the degradation results more consistent with the actual observation process.

[0171] Compared to traditional degradation methods that only employ fixed fuzziness or bicubic downsampling, the method of this invention constructs low-resolution degradation samples that better conform to the real remote sensing imaging mechanism through channel-independent radiometric fidelity mapping, multi-kernel family sampling mechanism, content-adaptive fuzzy parameter mapping, and physical constraint pruning based on the target sensor modulation transfer function. The generated degradation images are closer to the real sensor observation results in terms of band radiometric response, spatial fuzziness morphology, and frequency domain attenuation characteristics. It can better simulate the spatial fuzziness characteristics and frequency attenuation laws in real remote sensing sensor imaging, and has good stability and practicality. It is suitable for multispectral remote sensing image degradation modeling, training sample construction, and related remote sensing image processing applications, and can provide a more reliable data foundation for remote sensing image super-resolution reconstruction, image enhancement, and related downstream tasks.

Claims

1. A method for high-fidelity adaptive degradation of remote sensing image radiometrics controlled by modulation transfer function, characterized in that, Specifically, the following steps are included: Step 1: Obtain high-resolution multispectral remote sensing images from the target area or target dataset as input. The high-resolution multispectral remote sensing images are represented as follows: ,in: Represents the real number field. This indicates the input high-resolution multispectral remote sensing image. Indicates the number of spectral channels. Indicates the height of the input image. Indicates the width of the input image; Step 2: Perform channel normalization processing on each band of the high-resolution multispectral remote sensing image to obtain the normalized image. Step 3: Determine the blur kernel type based on the multi-kernel family sampling mechanism, and perform content-adaptive blur parameter mapping based on image content features to obtain blur kernel parameters; Step 4: Constrain and trim the fuzzy kernel parameters according to the imaging characteristics of the target sensor. First, determine the range of the equivalent point spread function parameters corresponding to the target sensor based on the publicly available spatial response information or modulation transfer function information of the target sensor. Then, restrict the parameters obtained by content adaptive fuzzy parameter mapping to this physical range. Step 5: Perform blur convolution degradation on the normalized image using the cropped blur kernel; Step 6: Perform downsampling on the image after blurring and convolution degradation, with a scaling factor of [value missing]. Downsampling yields low-resolution images; Step 7: Selectively add noise terms to the downsampled low-resolution image; Step 8, the normalized domain image after downsampling and optional noise overlay is denoted as ,right Perform channel-independent radiometric inverse normalization and constrain the inverse normalization results within the radiometric range of the corresponding band to obtain low-resolution degraded images. ,in: This indicates a low-resolution degraded image output. and These represent the height and width of the output image, respectively. , .

2. The method for high-fidelity adaptive degradation of remote sensing image radiometrics controlled by modulation transfer function according to claim 1, characterized in that, Step 2 is performed as follows: For the first input image For each band, the minimum radiation value is first calculated in the spatial dimension. and maximum radiation value , means as follows: In the formula: Indicates the input image number The spatial location of each band Pixel value at; Indicates the first Minimum radiation value for each band; Indicates the first Maximum radiation value of each band; Then for the first Each band undergoes independent channel normalization processing, as shown below: In the formula, Indicates the image after normalization. Each band in position Pixel value at; This is a preset minimum positive number to prevent the denominator from being zero; This indicates an operation to take the larger value.

3. The modulation transfer function controlled remote sensing image radiometric high-fidelity adaptive degradation method according to claim 2, characterized in that, The normalization process in Step 2 also includes outlier robustness processing, which employs a quantile-based robust mapping method, replacing the minimum radiation value with the preset values ​​corresponding to low and high quantiles, respectively. and maximum radiation value Normalization and denormalization are performed.

4. The modulation transfer function controlled remote sensing image radiometric high-fidelity adaptive degradation method according to claim 2, characterized in that, Step 3 involves the following steps: Step 3-1, Construct a multi-core family sampling mechanism: Preset fuzzy kernel family Recorded as: In the formula: Indicates candidate fuzzy kernel type I; Indicates candidate fuzzy kernel type II; Indicates candidate fuzzy kernel type III; The fuzzy kernel type is randomly sampled according to a mixed distribution, as follows: In the formula: , , They represent , , The sampling weights, and satisfying ; Indicates the type of fuzzy kernel obtained from sampling; This represents the specific fuzzy kernel instantiated based on the fuzzy kernel type; Step 3-2, Adaptive fuzzy parameter mapping based on image content features: Adopting a content-adaptive fuzzy parameter mapping strategy, the sampling range of fuzzy parameters is dynamically adjusted according to the texture complexity of the input image.

5. The modulation transfer function controlled remote sensing image radiometric high-fidelity adaptive degradation method according to claim 4, characterized in that, Step 3-1 Preset Fuzzy Kernel Family middle, Anisotropic Gaussian kernel, For the generalized radial exponent kernel, For a flat-topped radial core, the mathematical expression is as follows: In the formula: Indicates the coordinate position of the fuzzy kernel. ; Indicates the rotation angle A defined two-dimensional rotation matrix; and These represent the diffusion scale parameters along the main and secondary directions, respectively. is the normalization constant for the anisotropic Gaussian kernel; The Euclidean norm of a vector; Indicates the radial scale parameter; Indicates the tail shape parameters; , , These are the normalization constants for the anisotropic Gaussian kernel, the radial exponential kernel, and the flat-top radial kernel, respectively.

6. The modulation transfer function controlled remote sensing image radiometric high-fidelity adaptive degradation method according to claim 4, characterized in that, Step 3-2 The specific process is as follows: Step 3-2-1, input image Convert to grayscale reference image Grayscale reference image By using the preset channel set The pixel mean value of multiple channels in the image is obtained as follows: In the formula: Represents a grayscale reference image; Indicates a preset set of channels; Represents the preset channel set The number of channels in; Indicates the first input image One channel. A preferred embodiment uses a preset channel set. It is a set of visible light channels; Step 3-2-2: Calculate the grayscale reference images using the Sobel operator. Horizontal and vertical gradient components and And calculate the gradient magnitude of the entire image, as shown below: In the formula: Represents grayscale reference image In spatial location Gradient magnitude at; Represents grayscale reference image In spatial location The gradient component in the horizontal direction; Represents grayscale reference image In spatial location The gradient component in the vertical direction; Step 3-2-3: Take the average gradient magnitude of the entire image and use it as the texture score of the current sample. , means as follows: In the formula, This represents the texture score of the current input sample; Step 3-2-4: Normalize the texture scores Mapped to a blur parameter sampling interval to control the intensity range of the blur kernel, texture score The sampling interval of the fuzzy parameter satisfies the following monotonically decreasing relationship: In the formula: This represents the sampling interval of the fuzzy parameters corresponding to the current sample; This represents a mapping function from texture scores to the range of blur parameters. This indicates that the mapping function is monotonically decreasing with respect to the texture score.

7. The modulation transfer function controlled remote sensing image radiometric high-fidelity adaptive degradation method according to claim 6, characterized in that, Step 4 involves the following steps: Step 4-1: Construct parameter constraints based on the imaging characteristics of the target sensor: The equivalent point spread function of the target sensor is approximated by a two-dimensional anisotropic Gaussian function, as follows: In the formula: This represents the point spread function of the target sensor in the spatial domain. and Represents spatial coordinates; and These represent the diffusion parameters along the horizontal and vertical directions, respectively. Its corresponding frequency domain modulation transfer function (MTF) is approximately expressed as follows: In the formula: Represents the frequency domain modulation transfer function; and These represent the spatial frequency coordinates in the horizontal and vertical directions, respectively. The frequency domain modulation transfer function (MTF) is represented in a one-dimensional equivalent form as follows: In the formula: Represents frequency in one-dimensional space; This represents the one-dimensional equivalent diffusion parameter; Based on the MTF range of the target sensor at the Nyquist frequency, the one-dimensional equivalent diffusion parameters are obtained by inversion. The physically feasible range; Step 4-2, Parameter Trimming and Physical Consistency Constraints: When the target sensor mode is enabled, the parameters obtained by content-adaptive fuzzy mapping are intersected with the physical constraint range of the target sensor and then clipped, including at least: In the formula: and These represent the horizontal and vertical diffusion parameters obtained from the content-adaptive fuzzy mapping, respectively. , , , These represent the upper and lower bounds of the physical constraint range of the target sensor, respectively. This represents a clipping function used to restrict input parameters to a given range.

8. The modulation transfer function controlled remote sensing image radiometric high-fidelity adaptive degradation method according to claim 7, characterized in that, In Step 8, for Perform independent radial denormalization of the execution channel, as follows: In the formula: This represents the output pixel value after denormalization; Indicates the degenerate th term within the normalization domain. Each band in position Pixel value at; This indicates that the input value will be clipped to a range. Inside; and They represent the first The maximum and minimum radiation values ​​of the original radiation range of each band.

9. The method for high-fidelity adaptive degradation of remote sensing image radiometrics controlled by modulation transfer function according to claim 1, characterized in that, In Step 6, bicubic interpolation downsampling is used.

10. The modulation transfer function controlled remote sensing image radiometric high-fidelity adaptive degradation method according to claim 1, characterized in that, In Step 7, the superimposed noise term is at least one of Gaussian noise, Poisson noise, or a combination of noise corresponding to the noise characteristics of the target sensor.

Images