Two-dimensional feature selection verification method for similarity assessment of optical remote sensing images

Abstract

The present invention proposes a two-dimensional feature selection and verification method for optical remote sensing image similarity assessment, which belongs to the field of optical remote sensing detection image simulation model accuracy assessment. The method first selects the measured optical remote sensing data product of the area to be evaluated, extracts the detection parameters from the measured data, uses the optical remote sensing imaging simulation model to simulate the image of the extracted data, respectively performs two-dimensional Fourier transform, two-dimensional spatial frequency domain filtering and two-dimensional inverse Fourier transform processing on the measured image and the simulated image, extracts the DC component, low-frequency component and high-frequency component in the two images, compares and calculates the amplitude difference estimation, feature difference estimation and global difference estimation pixel by pixel, forms three types of estimated two-dimensional images, and uses the statistical histogram and cumulative probability density curve of the estimated image to evaluate the similarity of the two images. The present invention has a high sensitivity to the micro-change features of the image, and can have a good recognition and evaluation of the texture details of the image.

Description

Technical Field

[0001] The invention belongs to the field of accuracy assessment of optical remote sensing detection image simulation models, and in particular relates to a two-dimensional feature selection verification method for similarity assessment of optical remote sensing images. Background Art

[0002] The simulation model of optical remote sensing detection images is the basis for assisting the interpretation of optical remote sensing detection data, and plays an important role in understanding natural objective phenomena and the state of the earth's environment. The verification, validation and accreditation (VV&A) of the simulation model of optical remote sensing detection images is the basis for evaluating the accuracy of the simulation results in representing the real scene, which directly affects the reliability and scope of application of the model. The accuracy of the model is usually evaluated by comparing the similarity between the simulation and the measured data. The basic evaluation indicators include mean, variance, peak signal-to-noise ratio, mean square error, etc. However, each pixel unit in the optical remote sensing image contains important data about the characteristics and information of the ground object. The results obtained by using the overall evaluation index cannot reflect the similarity of each pixel unit in the image. Therefore, it is necessary to use pixel-by-pixel evaluation to more accurately compare the differences and similarities between images, thereby improving the accuracy and reliability of the evaluation.

[0003] The following evaluation methods exist in the prior art:

[0004] 1. Mean square error, peak signal-to-noise ratio, correlation coefficient and other energy error-based methods. The energy error-based method is to calculate the image distortion of the corresponding pixel units of the simulated image and the measured image one by one to obtain the difference of each pixel unit. Commonly used methods for calculating the difference include mean square error, peak signal-to-noise ratio and correlation coefficient, which are relatively simple in mathematical sense.

[0005] However, it has the following disadvantages: this type of method only compares the differences between the numerical values ​​of image pixel units, and does not take into account the differences in the human visual system's perception of image quality. Since human vision is more sensitive to contrast differences with lower spatial frequencies and less sensitive to brightness contrast differences, and the perception of a region is affected by its surrounding neighboring areas, the image similarity evaluation results in some cases are inconsistent with the subjective perception of human vision.

[0006] 2. Structural Similarity (SSIM) index. From the perspective of image composition, structural information is defined as an attribute that is independent of brightness and contrast and reflects the structure of objects in the scene. Distortion is modeled as a combination of three different factors: brightness, contrast, and structure. This results in an evaluation that considers the brightness, contrast, and structure of the image at the same time.

[0007] However, it has the following disadvantages: In order to obtain the distribution of evaluation results of pixel-by-pixel units, SSIM uses a sliding window to divide the image into blocks and calculate the average structural similarity of its window. The result will be affected by the size and shape of the window, and the result blurs the texture details of the image. Therefore, this method lacks sensitivity to the slight changes in the image. In addition, the SSIM indicator has limitations on use and cannot work effectively when the image has non-structural distortions such as displacement, scaling, and rotation.

[0008] 3. Two-dimensional data similarity evaluation method based on one-dimensional feature selective validation (FSV) method. The two-dimensional image data matrix is ​​split into row vector groups and column vector groups by row and column respectively, and the one-dimensional feature selective validation method is repeatedly used for evaluation. The weighted root mean square of the two row and column evaluation results of each pixel unit is calculated to obtain the distribution of similarity evaluation results of each pixel unit.

[0009] However, it has the following disadvantages: the one-dimensional feature selection verification method extracts the DC, low-frequency and high-frequency components based on the frequency domain energy of the data, and the energy extraction has a relative meaning. During the evaluation process, the evaluation results of each pixel unit may be affected by the outliers from the row vector or column vector, causing the final result to be contaminated to a certain extent. Therefore, this method has limited applicability for two-dimensional spatial domain image similarity evaluation. Summary of the invention

[0010] To solve the above technical problems, the present invention proposes a 2-Dimensional Feature Selective Validation (2D-FSV) method for similarity evaluation of optical remote sensing images. First, a measured image is obtained and its detection parameters are extracted as a basis for model parameter setting. A simulation image is generated using an optical remote sensing detection image simulation model. Second, two pre-processed remote sensing images are subjected to a 2D Fourier transform to obtain a spatial spectrum distribution. Then, filtering is performed in a 2D spatial frequency domain. After a 2D inverse Fourier transform, a DC component, a low-frequency component and a high-frequency component are extracted. Finally, the DC component and the low-frequency component are used to calculate the amplitude difference measurement (AmplitudeDifference Measurement, ADM) of each pixel unit, the low-frequency component and the high-frequency component are used to obtain the feature difference measurement (FeatureDifference Measurement, FDM), and the ADM and FDM are combined to obtain a global difference measurement (GlobalDifferenceMeasurement, GDM). In addition, multiple forms of evaluation results such as a statistical histogram, a cumulative probability density curve and the overall image similarity can also be obtained.

[0011] In order to achieve the above object, the technical solution adopted by the present invention is:

[0012] The two-dimensional feature selection verification method for optical remote sensing image similarity evaluation includes the following steps:

[0013] Step 1: Select optical remote sensing data products, obtain measured images and extract detection parameters, which include observation time, spectrum, and sensor parameters;

[0014] Step 2: Generate a simulated image using the optical remote sensing detection image simulation model, and set the model parameters to be the same as the detection parameters extracted from the measured image;

[0015] Step 3: Perform two-dimensional Fourier transform, two-dimensional spatial frequency domain filtering, and two-dimensional inverse Fourier transform on the measured image and the simulated image respectively, and extract the DC component, low-frequency component, and high-frequency component corresponding to the measured image and the simulated image respectively;

[0016] Step 4: Calculate a pixel-by-pixel unit distribution map of the comparative evaluation results of the measured image and the simulated image, wherein the DC component and the low-frequency component are calculated to obtain an amplitude difference estimate, the low-frequency component and the high-frequency component are calculated to obtain a feature difference estimate, and the amplitude difference estimate and the feature difference estimate are calculated to obtain a global difference estimate;

[0017] Step 5: Perform statistical processing on the amplitude difference estimation, feature difference estimation, and global difference estimation to obtain a statistical histogram, a cumulative probability density curve, and image similarity results.

[0018] The beneficial effects of the present invention are:

[0019] (1) The method of the present invention has good applicability. The present invention does not limit the source and form of the image to be evaluated. The image can come from the actual detection data of the sensor, or it can be the simulation calculation result of the mathematical or physical model. It can evaluate remote sensing images of any frequency band, and can evaluate radiation images, reflectivity images, digital images, etc.

[0020] (2) The method of the present invention has a high sensitivity to the subtle changes in the features of the image and can well identify and evaluate the texture details of the image. In addition, the evaluation index obtained by this method has an absolute standard for reference in terms of numerical values ​​and can be mapped to a qualitative evaluation level. The obtained evaluation results have actual physical meaning, which is difficult to achieve with most current evaluation methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 It is a flowchart of a two-dimensional feature selection and verification method for optical remote sensing image evaluation according to the present invention;

[0022] Figure 2A schematic diagram of a two-dimensional spatial frequency domain filter defined when extracting a DC component, a low-frequency component and a high frequency in a two-dimensional feature selection verification method;

[0023] Figure 3 (a) is a measured image obtained by the AGRI sensor of the Fengyun-4B satellite; Figure 3 (b) is a simulated image of the cloud-free Earth disk radiance in the 3.5-4.0 μm spectral band obtained using the optical remote sensing detection image simulation model;

[0024] Figure 4 (a) and Figure 4 (b) are the DC components extracted after performing 2D Fourier transform, 2D spatial frequency domain filtering and 2D inverse Fourier transform on the measured image and the simulated image, respectively;

[0025] Figure 5 (a) and Figure 5 (b) are low-frequency components extracted after two-dimensional Fourier transform, two-dimensional spatial frequency domain filtering, and two-dimensional inverse Fourier transform of the measured image and the simulated image, respectively;

[0026] Figure 6 (a) and Figure 6 (b) are high-frequency components extracted after two-dimensional Fourier transform, two-dimensional spatial frequency domain filtering, and two-dimensional inverse Fourier transform of the measured image and the simulated image, respectively;

[0027] FIG7(a), FIG7(b) and FIG7(c) are respectively the pixel-by-pixel amplitude difference estimation results, feature difference estimation results and global difference estimation results obtained by evaluating the measured image and the simulated image;

[0028] Figure 8 The statistical histogram and cumulative probability density curve are obtained by statistically processing the pixel-by-pixel results. DETAILED DESCRIPTION

[0029] The specific implementation of the present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0030] like Figure 1 As shown, the two-dimensional feature selection verification method for optical remote sensing image similarity simulation of the present invention comprises the following steps:

[0031] Step 1: Select optical remote sensing data products, obtain measured images and extract detection parameters, which include observation time, spectrum, and sensor parameters.

[0032] The file name of the optical sensor Level 1 data product usually includes the start and end time of the data observation. The orbital altitude and position of the sensor platform, spatial and temporal resolution and other information can be obtained from the public data of the satellite sensor. The spectral range and number can be obtained in the product appendix. The data product usually directly gives the DN value of the calibrated reflectivity or radiance after linear transformation. By performing simple linear operations with given multiplicative factors and additive factors, the calibrated reflectivity image and radiance image can be obtained.

[0033] Step 2: Generate a simulated image using the optical remote sensing detection image simulation model, and set the model parameters to be the same as the detection parameters extracted from the measured image.

[0034] The simulated image is generated by the optical remote sensing detection image simulation model. In order to obtain more accurate evaluation results, the measured image and the simulated image can be preprocessed as needed, which mainly includes three aspects. The first is the processing of invalid pixel unit values. It should be noted that the measured image is not always valid pixel unit values. In most cases, the part beyond the observation range is filled with invalid values, resulting in the calculated reflectivity or radiance being much higher or lower than the valid value, which expands the numerical variation range of the overall evaluation data. In the subsequent component extraction, its high brightness will cover the detail information. For this part, it is necessary to reassign it to 0 before the evaluation to improve its impact on the evaluation results. The corresponding simulated image can also be processed in the same way. The second is to perform denoising on the measured image. The measured image inevitably contains system noise and environmental noise, while the simulated image generally does not consider the impact of noise. Therefore, the measured image can be denoised first to reduce the impact of noise on the accuracy of the evaluation results. The third is the calibration of the measured image and the simulated image. By resampling, interpolating or intercepting the area of ​​interest, two images with completely aligned pixel units are obtained, and the position and spatial resolution of the images are calibrated. In addition, the solar spectrum radiance brightness values ​​of the two images can also be calibrated to avoid the influence of radiance deviation on the evaluation results.

[0035] Step 3: Perform two-dimensional Fourier transform, two-dimensional spatial frequency domain filtering and two-dimensional inverse Fourier transform on the measured image and the simulated image respectively to extract the DC component, low-frequency component and high-frequency component corresponding to the measured image and the simulated image respectively.

[0036] The two remote sensing images after preprocessing can be expressed in matrix form. Assuming that the measured image is , the simulation image is , the number of pixel units of the two images is , X(m,n) and Y(m,n) can be used to represent the image respectively and The pixel unit position is , that is, Row, No. The values ​​of the pixel units are listed in the table. It should be noted that all calls to the elements in the matrix in the present invention follow this form. The measured image and the simulated image are subjected to a two-dimensional Fourier transform. First, the two dimensions of the image matrix are padded with zeros. After the padded with zeros, the number of pixel units in the image is ,satisfy and , the number and method of zero padding will not affect the final result. A simple zero padding method can be expressed as:

[0037] ,

[0038] In the formula, and are the image matrices after zero padding of the measured image and the simulated image, and They are Measured image of and simulation images The matrix expression of for The total number of rows of the image after zero padding satisfies , the total number of image columns satisfies .

[0039] Performing a two-dimensional discrete Fourier transform on the two zero-filled images can obtain the two-dimensional spatial spectrum of the measured image. And the two-dimensional spatial spectrum of the simulated image , where the calculation of the value of each pixel unit can be expressed as:

[0040] ,

[0041] in, and They represent the measured image and the simulated image after zero padding. Row, No. The pixel cell value of the column, and Represent the two-dimensional spatial spectrum of the measured image And the two-dimensional spatial spectrum of the simulated image No. Row, No. Column pixel unit value, intermediate parameter , intermediate parameter , is an imaginary unit, and Respectively represent the total number of rows and columns of the image.

[0042] In the specific implementation of the two-dimensional Fourier transform, the two images can be transformed into the frequency domain using the two-dimensional fast Fourier transform. Since the spectrum obtained by directly processing the original data concentrates the high-frequency results in the center, and the farther away from the center, the lower the frequency, which is not conducive to the subsequent extraction of different frequency components, the spectrum needs to be centered so that the low-frequency results of the transformed spectrum are concentrated in the center and the frequencies away from the center are higher. Can be divided into , , , Four sub-arrays, corresponding to The upper left, upper right, lower left, and lower right positions of the simulated image are in two-dimensional space spectrum Can be divided into , , , Four sub-arrays, corresponding to The upper left, upper right, lower left, and lower right positions, namely:

[0043] ,

[0044] The number of rows and columns of the sub-matrix are represented by their respective subscripts. and Indicates rounding down and rounding up respectively.

[0045] Centering does not change the number of pixel units in the spectrum. The two-dimensional spatial spectrum of the measured image and the simulated image after centering and It can be expressed as:

[0046] ,

[0047] The two-dimensional frequency domain filter required for performing two-dimensional spatial frequency domain filtering on the image to extract the DC component, low-frequency component and high-frequency component is as follows: Figure 2 As shown, Figure 2 Only part of the two-dimensional frequency domain filter is shown, and the origin of the coordinate system corresponds to the center of the spectrum , the spatial spectrum distribution is divided into four regions (A), (B), (C), and (D) for different filtering processes. The number of pixel units is defined as The distance matrix , whose elements Indicates Row, No. The distance between the pixel unit of the column and the center of the spectrum The distance corresponding to the first Row, No. The distance between the pixel unit of the column and the center of the spectrum The distance is calculated as follows:

[0048] ,

[0049] Distance Matrix The largest element in for:

[0050] ,

[0051] The dividing point between the DC component and the low-frequency component is defined as , the dividing point between low-frequency component and high-frequency component is .for The selection of The value of is not sensitive, so it can be directly set to an empirical value ,and The selection needs to be calculated based on the total energy in the frequency domain. First, the two-dimensional spatial spectrum of the measured image and the simulated image after centering is calculated respectively. and Medium pixel unit and , from the center of the spectrum The distance is greater than The sum of the frequency domain data and ,Right now:

[0052] ,

[0053] The dividing point between low-frequency components and high-frequency components Select the two-dimensional spatial spectrum of the measured image after centering (or the two-dimensional spatial spectrum of the simulated image after centering ) The distance between the pixel unit and the center of the spectrum The distance from The sum of the frequency domain data accumulated at the beginning is just greater than (or ), that is:

[0054] ,

[0055] In the formula, and are the dividing points of low-frequency components and high-frequency components of the measured image and the simulated image calculated by the 40% principle. These two dividing points are usually different values, so they need to be unified into a common dividing point, that is:

[0056] ,

[0057] Determine the two component dividing points and After that, we can design a two-dimensional frequency domain filter and extract the frequency domain components. The filter for extracting the DC component is defined as matrix , among which Row, No. The value of the pixel cell of the column It can be expressed as:

[0058] ,

[0059] The DC component spectrum can be expressed as the two-dimensional spatial spectrum of the measured image and the simulated image after centering and the filter for extracting the DC component. The dot product of ,Right now:

[0060] ,

[0061] In the formula, It represents matrix dot multiplication, that is, matrix multiplication of elements, the result is still a matrix, and the number of pixel units does not change.

[0062] Then decentralization is performed, which can be understood as the inverse operation of centralization. The DC component spectrum Can be divided into , , , Four sub-arrays, corresponding to The upper left, upper right, lower left, and lower right positions. Can be divided into , , , Four sub-arrays, corresponding to The positions of the upper left, upper right, lower left, and lower right in the middle, the number of rows and columns of the subarray are represented by their respective subscripts, and Indicates rounding down and rounding up respectively.

[0063] ,

[0064] The DC component spectra of the measured image and the simulated image after decentralization are and :

[0065] ,

[0066] Decentered DC component spectrum of measured and simulated images and Perform two-dimensional discrete inverse Fourier transform to obtain the time domain distribution results of the DC component of the measured image and the simulated image and , the calculation of the value of each pixel unit can be expressed as:

[0067] ,

[0068] in, and The decentered DC component spectra of the measured image and the simulated image are represented by Row, No. The pixel cell value of the column, and They represent the time domain distribution of the DC component of the measured image after two-dimensional inverse Fourier transform. And the time domain distribution of DC component of the simulated image No. Row, No. Column pixel unit value, intermediate parameter , intermediate parameter , is an imaginary unit, and They respectively represent the total number of rows and the total number of columns of the matrix. In the specific implementation of the two-dimensional inverse Fourier transform, the two images can be transformed into the time domain using a two-dimensional inverse fast Fourier transform.

[0069] Since the original measured and simulated images are padded with zeros during the two-dimensional Fourier transform, the number of pixel units of the DC component obtained at this time is , the number of image pixel units that need to be restored to before zero padding , that is, the DC component of the final measured image and simulated image and They are and The middle position is the upper left, and the number of pixel units is The subarray of , namely:

[0070] D C ′ X = [ ( DC X ) M × N A M × N B M × N C M × N ] U × V D C ′ Y = [ ( DC Y ) M × N A ′ M × N B ′ M × N C ′ M × N ] U × V ,

[0071] Among them, A, B, and C represent The middle position is the upper right, lower left, and lower right subarrays, and the number of pixel units is , A', B', C' are The middle position is the upper right, lower left, and lower right subarrays, and the number of pixel units is , thus completing the extraction of DC components of measured images and simulated images.

[0072] The filter for extracting low-frequency components is defined as matrix , among which Row, No. The value of the pixel unit It can be expressed as:

[0073] ,

[0074] The low-frequency component spectrum can be expressed as the two-dimensional spatial spectrum of the measured image and the simulated image after centering and the filter for extracting the low-frequency component The dot product of ,Right now:

[0075] ,

[0076] In the formula, It represents matrix dot multiplication, that is, matrix multiplication of elements, the result is still a matrix, and the number of pixel units does not change.

[0077] The final low-frequency component can be obtained through decentralization, two-dimensional inverse Fourier transform and subarray extraction. and , the number of pixel units is .

[0078] The filter for extracting high-frequency components is defined as matrix , among which Row, No. The value of the pixel cell of the column It can be expressed as:

[0079]

[0080] It is worth noting that the filter that extracts the low-frequency components and filters to extract high frequency components They have a complementary relationship, and the sum of the pixel unit values ​​at the same position is 1.

[0081] The high-frequency component spectrum can be expressed as the two-dimensional spatial spectrum of the measured image and the simulated image after centering. and and filters to extract high frequency components The number of pixel units is ,Right now:

[0082] ,

[0083] In the formula, It represents matrix dot multiplication, that is, matrix multiplication of elements, the result is still a matrix, and the number of pixel units does not change.

[0084] The final high-frequency component can be obtained through decentralization, two-dimensional inverse Fourier transform and subarray extraction. and , the number of pixel units is .

[0085] Step 4: Calculate the pixel-by-pixel unit distribution map of the comparative evaluation results of the measured image and the simulated image, wherein the DC component and the low-frequency component are calculated to obtain the amplitude difference estimate (ADM), the low-frequency component and the high-frequency component are calculated to obtain the feature difference estimate (FDM), and the amplitude difference estimate and the feature difference estimate are calculated to obtain the global difference estimate (GDM).

[0086] 2D-FSV uses three evaluation indicators to evaluate the evaluation object with different focuses. Among them, ADM reflects the degree of difference in image amplitude, FDM reflects the difference in change trends between image data, and respectively reflects the characteristic differences between the slow-changing components and the fast-changing components of the data. GDM is obtained by combining ADM and FDM, reflecting the overall difference in amplitude and change characteristics of the data, and is a global comprehensive judgment on the similarity between images.

[0087] definition is the pixel-by-pixel evaluation result, and its dimension is the same as the measured image. , then the evaluation result of pixel-by-pixel ADM can be expressed as a matrix , the value of its pixel unit The calculation method is:

[0088] ,

[0089] in, express The matrix Row, No. The pixel unit of the column, It represents the absolute value difference between the low-frequency component of the measured image and the low-frequency component of the simulated image on a pixel-by-pixel basis. Represents the mean of the sum of the absolute values ​​of the low-frequency components of the measured image and the simulated image per pixel, It represents the absolute value difference between the DC component of the measured image and the DC component of the simulated image on a pixel-by-pixel basis. It represents the mean of the absolute value of the pixel-by-pixel sum of the DC component and low-frequency component of the measured image and the DC component and low-frequency component of the simulated image. exp represents the exponential function, that is:

[0090] ,

[0091] It can be concluded from the calculation method of ADM that the ADM assessment of the degree of difference in amplitude is obtained by calculating the DC component and the low-frequency component. The point-by-point amplitude difference is reflected by the degree of change of the low-frequency component amplitude difference of each pixel unit relative to the overall average amplitude. At the same time, in order to consider the influence of the DC component on the amplitude difference, the degree of change of the DC component amplitude difference of each pixel unit relative to the overall average amplitude is calculated, and its weight in the ADM calculation is defined in the form of an exponent. When the DC component difference is small, the weight of this item is small. As the DC component difference increases, the weight of this item increases nonlinearly and synchronously, resulting in a nonlinear increase in the final ADM evaluation result. At this time, the evaluation results of some pixel units with larger values ​​may mask the evaluation of the difference in image details. Therefore, when the DC difference between the two images is significantly large, the item containing the DC component can be discarded in the calculation of ADM.

[0092] The evaluation result of pixel-by-pixel FDM can be expressed as a matrix , the value of its pixel unit It is obtained by calculating the derivatives of the low-frequency and high-frequency components of the data, namely:

[0093] ,

[0094] in, express The matrix Row, No. The pixel unit of the column, It indicates the difference between the first-order derivative change characteristics of the low-frequency component and the overall average value, and the weight assigned is 1 / 2. It indicates the difference between the first-order derivative change characteristics of the high-frequency component and the overall average value, and the weight assigned is 1 / 6. It indicates the difference between the second-order derivative change characteristics of the high-frequency component and the overall average value, and the weight assigned is 1 / 7.2. represents the first-order derivative operation, the superscript represents the second-order derivative operation, that is:

[0095] ,

[0096] According to the mathematical meaning of the derivative, the first-order derivative describes the magnitude of the data change rate, and the second-order derivative describes the high-order change characteristics of the data, including the distribution of peaks and valleys. The calculation forms of the three parts that make up the FDM are similar, all of which represent the degree of difference of a certain feature of the image relative to the overall average value, but the three have different proportions in the calculation of the overall FDM, namely 1 / 2, 1 / 6, and 1 / 7.2, respectively. That is, it is believed that the difference in low-frequency component changes contributes more to the overall evaluation results than the difference in high-frequency component changes, and the low-order change difference contributes more to the high-order change difference. This proportion can also be flexibly adjusted according to the specific application situation.

[0097] In practical applications, the final evaluation result is not sensitive to the calculation method of the matrix derivative. Here is a general calculation process of the matrix derivative, which can be obtained by combining the definition of the central difference and the gradient. The total number of pixel unit rows is set to The total number of columns is The matrix R, whose Row, No. The first-order derivative calculation of the pixel unit of the column is expressed as:

[0098] ,

[0099] Substituting the calculation results of the first-order derivative of each pixel unit into the following process, the second-order derivative of each pixel unit can be obtained, which is expressed as:

[0100] ,

[0101] in:

[0102] ,

[0103] The evaluation result of pixel-by-pixel GDM can be expressed as a matrix , its pixel unit It is calculated comprehensively from the pixel units of ADM and FDM, namely:

[0104] ,

[0105] In the formula, and are the weights of ADM and FDM respectively, satisfying the condition + =1, generally, ADM and FDM can be given the same weight, that is, =0.5, =0.5. In addition, it can be changed according to the actual evaluation focus, or it can be set according to the confidence level of the two estimation results.

[0106] Step 5: Perform statistical processing on the amplitude difference estimation, feature difference estimation, and global difference estimation to obtain a statistical histogram, a cumulative probability density curve, and image similarity results.

[0107] According to the evaluation result classification mapping table of the two-dimensional feature selection verification method in Table 1, the quantitative evaluation results can be mapped to qualitative descriptions to evaluate the degree of consistency between the data. The final evaluation results can be described in natural language as "excellent", "very good", "good", "average", "poor" and "very poor". Data statistics and qualitative mapping are also applicable to the three indicators of 2D-FSV. Evaluation results of pixel-by-pixel units The pixel unit values ​​in the effective area are classified and counted according to the qualitative evaluation mapping table, and the proportion of the data volume belonging to each category to the total data volume is combined into a probability density histogram of six boxes. , expressed as:

[0108] ,

[0109] Among them, p represents the probability that the evaluation result of each pixel unit in the effective area falls into each level, and the subscript represents its level.

[0110] Table 1

[0111]

[0112] The cumulative probability density curve is obtained by plotting the probability density histogram of the six bins. It is obtained by cumulative calculation. The single value result obtained by statistically analyzing the pixel-by-pixel unit evaluation results can describe the degree of difference of the data as a whole, where the mean represents the average difference of the data, the median represents the difference in the middle level of the data, and the standard deviation represents the concentration of the data difference distribution. The data similarity assignment in Table 1 can characterize how many areas of data at each level match, and the "excellent" similarity is defined as 100%, "extremely poor" is defined as 0%, and "average" has a qualified meaning and is defined as 60%. The remaining "very good", "good", and "poor" similarities are defined as 90%, 75%, and 30%, respectively. By processing the pixel-by-pixel evaluation results of the effective area, an overall data similarity result can be obtained. ,Right now:

[0113] . Example

[0114] The following is a specific example of the process of processing and evaluating data by the 2D-FSV method to further illustrate the present invention, where the measured image is a sensor image detected by the Advanced Geostationary Radiation Imager (AGRI) on the Fengyun-4B satellite (FY-4B), and the simulated image is generated by the optical remote sensing detection image simulation model. The Fengyun-4B satellite is a geostationary orbit satellite with an orbit altitude of about 35786km and a subsatellite point located at the equator, 133 km east longitude. o Position, a full disk scan can be performed in 15 minutes, the number of channels is 15, and the embodiment uses the medium-wave infrared wavelength 3.5~4.0μm spectrum, corresponding to channel 7 of the AGRI sensor.

[0115] Step 1: Select optical remote sensing data products, obtain measured images and extract detection parameters, which include observation time, spectrum, and sensor parameters.

[0116] The start time of the AGRI sensor acquisition of measured images by the Fengyun-4B satellite is 5:30:00 on January 11, 2024, and the end time is 5:44:59 on January 11, 2024, at the orbital position of 133 E. o The geostationary orbit above the Earth has a detection spectrum of 3.5-4.0 μm, a spatial resolution of about 2 km, and a scene image size of 5496 × 5496 pixels. The preprocessed remote sensing image is shown in Figure 3(a). The unit is W / m 2 / sr / μm.

[0117] Step 2: Generate a simulated image using the optical remote sensing detection image simulation model, and set the model parameters to be the same as the detection parameters extracted from the measured image.

[0118] The model parameters are set to be the same as the extraction parameters of the measured image. The simulated image of the cloud-free Earth disk radiance in the 3.5-4.0 μm spectral band obtained by using the optical remote sensing detection image simulation model is shown in Figure 3(b) after preprocessing. The unit is the same as that of the measured image, which is W / m 2 / sr / μm, the simulated image is a cloud-free radiation image, while the measured radiation image inevitably contains clouds, and the clouds are randomly covered and have a large coverage area. It is difficult to obtain good consistency between the two images, and the visual evaluation results are poor.

[0119] Step 3: Perform two-dimensional Fourier transform, two-dimensional spatial frequency domain filtering and two-dimensional inverse Fourier transform on the measured image and the simulated image respectively to extract the DC component, low-frequency component and high-frequency component corresponding to the measured image and the simulated image respectively.

[0120] Figures 4(a) and 4(b) are the DC components extracted by processing the measured image and the simulated image, respectively, which can reflect the change of the overall amplitude of the image; Figures 5(a) and 5(b) are the low-frequency components extracted by processing the measured image and the simulated image, respectively, which describe the slow-changing characteristics of the image in amplitude; Figures 6(a) and 6(b) are the high-frequency components extracted by processing the measured image and the simulated image, respectively, which describe more detailed change characteristics. Overall, the extracted frequency components can reflect the change characteristics of the image at different frequencies. Since the radiance of the cloud layer in the medium-wave infrared radiation image is higher than that of the ground surface, the radiance of the measured image containing clouds is generally higher than that of the simulated image. The difference can be clearly seen from the comparison of the DC components of the two. According to the previous analysis of ADM, the DC component with a large difference will occupy a large proportion in the evaluation result of ADM, which will cover the evaluation of detailed features and directly affect the results of GDM. Therefore, the evaluation result in this case can ignore the DC component, that is, the item containing the DC component can be discarded when calculating ADM.

[0121] Step 4: Calculate the pixel-by-pixel distribution map of the comparative evaluation results of the measured image and the simulated image, wherein the DC component and the low-frequency component are calculated to obtain the amplitude difference estimation, the low-frequency component and the high-frequency component are calculated to obtain the feature difference estimation, and the amplitude difference estimation and the feature difference estimation are calculated to obtain the global difference estimation.

[0122] Figures 7(a), 7(b) and 7(c) are the pixel-by-pixel ADM, FDM and GDM results obtained by evaluating the measured image and the simulated image, respectively. The dark areas in the 2D-FSV results indicate that the pixel units have a higher similarity, while the bright areas indicate a lower similarity. From the results, ADM and FDM reflect the difference distribution of the image in amplitude and detail, respectively, and GDM reflects the overall difference distribution of the image. Its light and dark changes can better reflect its change differences. It can be observed that 2D-FSV is very sensitive to detail changes, so it can identify more differences in texture details. In addition, the areas with larger differences generally correspond to cloud-covered areas, while the cloud-free areas have a higher similarity, which is consistent with the results of visual evaluation. This evaluation result related to cloud distribution also illustrates the possibility of applying this method to cloud detection.

[0123] Step 5: Perform statistical processing on the amplitude difference estimation, feature difference estimation, and global difference estimation to obtain a statistical histogram, a cumulative probability density curve, and image similarity results.

[0124] Figure 8The statistical histogram and cumulative probability density curve are obtained by statistically processing the pixel-by-pixel evaluation results of the three indicators. From the distribution histogram, the evaluation results tend to be "poor", but the distribution is relatively scattered, showing a low confidence level; from the cumulative curve, the pixel-by-pixel rating results of the three indicators at "general" and below are all over 50%, and the GDM results are dominated by the less consistent FDM results. The overall image similarity is poor, but in fact this result is still consistent with the visual and expected results. The calculated similarity results of the overall image are 63.15% for ADM, 36.82% for FDM, and 37.40% for GDM. The results are close to the visual evaluation results, and the corresponding rating is "poor".

[0125] The present invention can also adopt alternative solutions that can also achieve the purpose of the invention: preprocessing the image is only to obtain a more accurate evaluation result. The preprocessing process is not limited to the invalid value filling, noise reduction and calibration described in the present invention. The image data can be transformed and processed as needed to achieve the similarity evaluation of optical remote sensing images; the use of wavelet transform ideas or adjustment of the transition band and form of the two-dimensional spatial frequency domain filter can achieve the purpose of extracting different frequency components and achieve the similarity evaluation of optical remote sensing images; the method described in the present invention has many parameters that can be flexibly adjusted, such as the dividing point between the low-frequency component and the high-frequency component. The 40% principle used in the selection, the proportion of the three parts in the FDM calculation, and the weighted weights of ADM and FDM in the GDM calculation can be adjusted according to the specific application requirements and focus to achieve the similarity evaluation of optical remote sensing images.

[0126] The specific embodiments described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A two-dimensional feature selection verification method for similarity assessment of optical remote sensing images, characterized in that: The steps include: Step 1: Select optical remote sensing data products, obtain measured images and extract detection parameters, which include observation time, spectrum, and sensor parameters; Step 2: Generate a simulated image using the optical remote sensing detection image simulation model, and set the model parameters to be the same as the detection parameters extracted from the measured image; Step 3: Perform two-dimensional Fourier transform, two-dimensional spatial frequency domain filtering, and two-dimensional inverse Fourier transform on the measured image and the simulated image respectively, and extract the DC component, low-frequency component, and high-frequency component corresponding to the measured image and the simulated image respectively; Step 4: Calculate a pixel-by-pixel unit distribution map of the comparative evaluation results of the measured image and the simulated image, wherein the DC component and the low-frequency component are calculated to obtain an amplitude difference estimate, the low-frequency component and the high-frequency component are calculated to obtain a feature difference estimate, and the amplitude difference estimate and the feature difference estimate are calculated to obtain a global difference estimate; Step 5: Perform statistical processing on the amplitude difference estimation, feature difference estimation, and global difference estimation to obtain a statistical histogram, a cumulative probability density curve, and image similarity results; The step 3 comprises: The measured image and the simulated image are transformed into the frequency domain by two-dimensional Fourier transform, and different frequency components are extracted by two-dimensional spatial frequency domain filter. Assume that the number of pixel units of the measured image and the simulated image is , the origin of the coordinate system of the two-dimensional frequency domain filter corresponds to the center of the spectrum , define the number of pixel units as The distance matrix ,element Indicates Row, No. The distance between the pixel unit of the column and the center of the spectrum Distance: ， The dividing point between the DC component and the low-frequency component is defined as , the dividing point between low-frequency component and high-frequency component is , filter to extract the DC component Is the size of The matrix of Row, No. The value of the pixel cell of the column It is expressed as: ， The DC component spectrum is represented as the two-dimensional spatial spectrum after the centering of the measured image and the simulated image. and and filters to extract DC components The dot product of ,Right now: ， In the formula, It means matrix dot multiplication, that is, matrix multiplication of elements, the result is still a matrix, and the number of pixel units does not change. and The DC component spectra of the measured image and the simulated image are extracted respectively. The DC components of the final measured image and the simulated image are obtained by decentralization, two-dimensional inverse Fourier transform and subarray extraction. and , the number of pixel units is ; Define a filter to extract low frequency components Is the size of The matrix of Row, No. The value of the pixel cell of the column It is expressed as: ， The low-frequency component spectrum is represented as the two-dimensional spatial spectrum of the measured image and the simulated image after centering. and and filters to extract low-frequency components The dot product of ,Right now: ， In the formula, and The low-frequency component spectra of the measured image and the simulated image are extracted respectively. The final low-frequency components of the measured image and the simulated image are obtained by decentralization, two-dimensional inverse Fourier transform and subarray extraction. and , the number of pixel units is ; Define a filter to extract high frequency components Is the size of The matrix of Row, No. The value of the pixel cell of the column It is expressed as: ， The high-frequency component spectrum is represented as the two-dimensional spatial spectrum of the measured image and the simulated image after centering. and and filters to extract high frequency components The number of pixel units is ,Right now: ， In the formula, and The high-frequency component spectra of the measured image and the simulated image are extracted respectively. The high-frequency components of the final measured image and the simulated image are obtained by decentralization, two-dimensional inverse Fourier transform and subarray extraction. and , the number of pixel units is .

2. The two-dimensional feature selection verification method for optical remote sensing image similarity evaluation according to claim 1, characterized in that: The step 4 comprises: The amplitude difference estimation reflects the difference in amplitude of the image data. The evaluation result of the pixel-by-pixel amplitude difference estimation is expressed as matrix , the value of its pixel unit The calculation method is: ， in, express The matrix Row, No. The pixel unit of the column, It represents the absolute value difference between the low-frequency component of the measured image and the low-frequency component of the simulated image on a pixel-by-pixel basis. Represents the mean of the sum of the absolute values ​​of the low-frequency components of the measured image and the simulated image per pixel, It represents the absolute value difference between the DC component of the measured image and the DC component of the simulated image on a pixel-by-pixel basis. It represents the mean of the absolute value of the pixel-by-pixel sum of the DC component and low-frequency component of the measured image and the DC component and low-frequency component of the simulated image. exp represents the exponential function, that is: ， Feature difference estimation reflects the difference in the changing trend between image data. The evaluation result of pixel-by-pixel feature difference estimation is expressed as matrix , the value of its pixel unit The calculation method is: ， in, express The matrix Row, No. The pixel unit of the column, It indicates the difference between the first-order derivative change characteristics of the low-frequency component and the overall average value, and the weight assigned is 1 / 2. It indicates the difference between the first-order derivative change characteristics of the high-frequency component and the overall average value, and the weight assigned is 1 / 6. It indicates the difference between the second-order derivative change characteristics of the high-frequency component and the overall average value, and the weight assigned is 1 / 7.

2. represents the first-order derivative operation, the superscript represents the second-order derivative operation, that is: ， The global difference estimation reflects the overall difference of image data in amplitude and change trend. The evaluation result of pixel-by-pixel global difference estimation is expressed as matrix , its pixel unit It is calculated by the pixel units of amplitude difference estimation and feature difference estimation, that is: ， In the formula, and are the weights of amplitude difference estimation and feature difference estimation, respectively, satisfying + =1.

Images