A method for establishing radiation benchmark and radiation correction of high-resolution remote sensing images
Radiation invariant pixels are extracted through invariant detection and timing analysis methods, and relative radiation correction is performed in combination with Chi-square distribution statistics, which solves the problem of radiation inconsistency in remote sensing images, realizes high-precision radiation consistency correction and automated processing, and improves remote sensing Image quality evaluation and production efficiency.
Patent Information
- Application Number
- CN202111479931.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-12-06
- Publication Date
- 2024-12-24
- Estimated Expiration
- 2041-12-06
AI Technical Summary
In the prior art, remote sensing images have radiation inconsistencies due to differences in imaging sensors and constellations, which affects the accuracy of land use classification and natural resource monitoring. Traditional radiation correction methods are greatly affected by human subjectiveness, resulting in inconsistent radiation quality.
Invariant detection and timing analysis methods are used to extract radiation invariant pixels as radiation reference, and relative radiation correction is performed through chi-square distribution statistics to reduce artificial intervention and improve the radiation consistency of multi-source, multi-phase images.
High-precision radiation consistency correction is achieved, reducing human subjective influence, improving the radiation consistency and correction accuracy of multi-source multi-phase images, forming an automated and standardized radiation reference, and improving the quality evaluation of digital orthophoto images and Productivity.
Smart Images
Figure CN114359066B_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the field of remote sensing mapping, and in particular relates to a method for establishing a radiation benchmark and correcting radiation for a high-resolution remote sensing image. Background Art
[0002] With the continuous improvement of earth observation remote sensing technology, a large amount of remote sensing satellite image data can now be obtained, providing data support for scientific research, natural resource monitoring, national land space planning, disaster monitoring, etc. Due to the different imaging sensors and satellite constellations of different satellites, there are differences in the re-entry period, coverage, resolution, and spectral response coefficient, resulting in serious radiation differences between different images of the same object, that is, different spectra for the same object and the same spectrum for different objects. This problem seriously affects the engineering application of remote sensing images, especially when land use classification is difficult to accurately distinguish the types of objects. In addition, when using multi-temporal images to monitor natural resources, there are a large number of pseudo-changes, resulting in poor change detection accuracy and difficulty in engineering application. Therefore, the radiation consistency of cross-sensor images determines the usability of remote sensing images in engineering. In order to reduce the radiation differences between multi-source images, it is very important to establish a radiation benchmark to improve the accuracy of radiation correction.
[0003] Radiation correction is mainly divided into absolute correction and relative correction methods. Absolute radiation correction converts meaningless image values into surface reflectivity, which can reflect the physical properties of the surface. Usually, due to various errors in absolute radiation correction, there are still radiation differences after absolute correction, such as atmospheric model errors, imaging sensor coefficient errors, satellite constellation errors, etc. In order to further improve the radiation consistency, the traditional relative radiation method will select two images to manually sample the standard radiation values of typical objects for radiation correction. Traditional methods are greatly affected by human subjectivity and the selection of reference images. Most of them only meet the user's visual color consistency. In essence, they are relative radiation corrections under arbitrary radiation references (manually or through uniform light algorithms). There are many human interventions during the process, and the reference images are not unified. Differences in operators and working time will affect the uniform light effect. Therefore, the radiation quality of traditional methods varies, resulting in inconsistent standards.
[0004] In summary, establishing a ground object radiation benchmark and using geographic coordinate matching can provide a reference benchmark for relative radiation correction, provide a unified radiation benchmark, and realize automated radiation correction, reduce human subjective judgment, and improve the radiation consistency of multi-source and multi-temporal images.
[0005] Therefore, the present invention proposes a method for establishing a radiation benchmark and radiation correction for high-resolution remote sensing images. The invariant detection and time series analysis methods are used to extract invariant and deterministic pixels as radiation benchmarks, and chi-square distribution statistics are introduced to perform relative radiation correction. Summary of the invention
[0006] The invention establishes a radiation benchmark suitable for multi-phase and multi-source remote sensing image radiation correction, improves radiation consistency, and automates correction to reduce human subjective influence. To achieve the above functions, the invention solves the following technical problems: selection of radiation-invariant pixels and radiation-deterministic pixels; relative radiation correction optimization.
[0007] The present invention proposes a method for establishing a radiation benchmark and correcting radiation for high-resolution remote sensing images, comprising the following steps:
[0008] Step 1: determine the study area and obtain the homologous image sequence of the area in recent years as the image for constructing the radiation benchmark of the area;
[0009] Step 2: Based on the deep learning-based invariant detection method, the invariant regions of the homologous time series images are selected;
[0010] Step 3, performing a time series analysis on the pixels in the unchanged area to determine the radiation unchanged pixels and the radiation deterministic pixels, which are used as radiation benchmarks for radiation correction;
[0011] Step 4: First, preprocess the image to be corrected, then construct a chi-square distribution with the radiation benchmark determined in steps 1-3, perform statistical linear regression, remove abnormal pixels, and improve the relative radiation correction accuracy.
[0012] Furthermore, in step 1, the homologous image sequence of the study area for the past three years or more is taken as the image for constructing the radiation benchmark of the area. Specifically, the Sentinel-2 image data is used as the homologous image sequence. The homologous image meets the following conditions:
[0013] (1) To reduce the radiation differences caused by the systematic errors of the imaging equipment, images from the same satellite and the same sensor were selected;
[0014] (2) The study area should contain image data for three years or more, and ensure that one image can be obtained every quarter or month.
[0015] Furthermore, the specific implementation of step 2 is as follows;
[0016] Pairing time-series images 1 and I 2 Input into the invariant detection model based on deep learning respectively, and calculate and Conduct training. The specific training process is as follows:
[0017] Step 2.1, randomly crop the input image as training data, input it into Net1 and Net2 respectively, Net1 and Net2 are Swin-transformer, ResNet or VGG models, and then perform softmax calculation to obtain probability distribution; first, image I 1Input Net1 and Net2 networks for training to obtain image I 1 The probability distributions of the training results P1-1 and P1-2 are calculated, and the image I 1 of
[0018] Step 2.2: Image I 2 Input Net1 and Net2 networks for training to obtain image I 2 The probability distributions P2-1 and P2-2 of the training results are calculated, and the image I 2 of
[0019] Step 2.3, use the following formula to calculate the loss of the two images contrast , and then calculate Use gradient descent to update Net1 parameters, and use sliding average to update Net2 parameters;
[0020]
[0021] in, represents the mean function, h represents the feature encoding function based on conditional probability density, and Represents image I 1 The output value of Net1 and Net2, j represents The jth sample drawn from , k represents the number of samples drawn.
[0022] Step 2.4, finally use the trained model to detect the time series images, obtain the probability distribution map, and then use the binary classification method to obtain the change and no change categories, where the binary classification method uses the Ostu or K-means clustering algorithm.
[0023] Furthermore, the specific implementation of step 3 is as follows;
[0024] Step 3.1: Convert the image acquisition time into annual days as the time series independent variable d, and the pixel radiation value as the dependent variable V rc ;
[0025]
[0026] Among them, V rc represents the pixel value of the rth row and cth column of the time series image, d represents the annual accumulation day of the time series image, a 1 , a 2 , a 3 , a 4 , a 5 represents the time series model coefficient;
[0027] In step 3.2, using three times the curve regression error as the threshold, the pixels with three times the error of the fitting time series model curve are automatically and iteratively removed. The pixels left after removal are suitable for constructing the radiation benchmark for subsequent radiation correction.
[0028] Furthermore, the specific implementation of step 4 is as follows;
[0029] Step 4.1, firstly, the image to be corrected is subjected to absolute radiometric calibration and atmospheric correction, and then geometrically corrected with the radiometric reference pixels established by steps 1-3 to achieve sub-pixel matching accuracy of geographic coordinates, and then matched with the geographic coordinates between the radiometric reference and the image to be corrected;
[0030] Step 4.2, combine the matched pixels with linear regression to construct chi-square distribution and remove some unstable pixels;
[0031]
[0032]
[0033] Among them, a j and b j represents the relative radiation correction coefficient of the j-band image, and represents the radiation value of the i-th pixel of the j-band of the image to be corrected and the j-band of the reference image, n represents the number of pixels involved in the calculation, T represents the chi-square distribution constructed using the root mean square error of linear regression, and χ belongs to the symbol of the chi-square distribution;
[0034] In step 4.3, linear regression is performed on the pixels filtered by the chi-square distribution to determine the linear model of radiation correction. Finally, the radiation value of the image to be corrected is substituted into the linear model for calculation to perform radiation correction.
[0035] The present invention has the following advantages: 1. Adopting the invariant detection method, the radiation benchmark range is determined and selected, and the efficiency of establishing the radiation benchmark is improved; 2. The radiation benchmark range is analyzed pixel by pixel in time series, and the radiation invariant pixels and radiation determined pixels are extracted with high precision; 3. The chi-square distribution is constructed in combination with the image to be corrected, which effectively improves the accuracy of radiation correction. 4. The radiation benchmark standardizes the remote sensing satellite image, and provides an automated and standardized quantitative evaluation for the quality evaluation of digital orthophoto results. Subsequently, a high-performance digital orthophoto quality evaluation can be formed, and the production efficiency of digital orthophotos can be improved; 5. The establishment of the entire radiation benchmark and the radiation correction are both automated, without manual intervention, eliminating the subjective influence of human beings. BRIEF DESCRIPTION OF THE DRAWINGS
[0036] Figure 1 The figure is a flow chart of the method of the present invention.
[0037] Figure 2Schematic diagram of the invariant detection method.
[0038] Figure 3 The graphs are the results of timing analysis. (a), (b), (c), and (d) represent the results of blue, green, red, and near-infrared band processing, respectively.
[0039] Figure 4 The results of chi-square distribution screening and radiation correction are shown in Figure 2. (a), (b), (c), and (d) represent the results of blue, green, red, and near-infrared band processing, respectively. DETAILED DESCRIPTION
[0040] In order to make the purpose, technical solution and advantages of the present application more clear, the present application is further described in detail below in conjunction with specific examples. It should be understood that the specific examples described here are only used to explain the present application and are not used to limit the present application.
[0041] Combine the following Figures 1 to 4 The specific implementation mode of the present invention is a method for establishing a radiation benchmark and correcting radiation for a high-resolution remote sensing image, comprising the following steps:
[0042] Step 1: Determine the study area and obtain homologous image sequences of the area for the past three years or more as images for constructing the radiation benchmark of the area. This example will use Sentinel-2 image data as the homologous image sequence, and the homologous image should meet the following conditions:
[0043] 1. To reduce the radiation differences caused by the system errors of the imaging equipment, select images from the same satellite and the same sensor;
[0044] 2. The study area should contain image data for the past three years or more, and ensure that one image can be obtained every quarter or month;
[0045] Step 2: Figure 2 The figure shows an invariant detection model based on deep learning, which pairs the time series images. 1 and I 2 Enter this model separately and calculate and Conduct training and detect the unchanged area; the specific training process is as follows:
[0046] Step 2.1: Randomly crop the input image as training data and input it into the twin networks Net1 and Net2 respectively (Net1 and Net2 belong to two twin networks with the same parameters, but the difference between Net2 and Net1 is that Net2 does not back-propagate to update the coefficients, but uses the sliding average (ema) to directly assign Net1 parameters. Net1 and Net2 can be Swin-transformer, ResNet, VGG and other models). After softmax calculation, the softmax function is generally used as the last layer of the neural network, accepting the input value from the previous layer of the network and then converting it into probability. First, the image I 1 Input Net1 and Net2 networks for training to obtain image I 1 The probability distributions of the training results P1-1 and P1-2 are calculated, and the image I 1 of
[0047] Step 2.2: Similarly, image I 2 Input Net1 and Net2 networks for training to obtain image I 2 The probability distributions P2-1 and P2-2 of the training results are calculated, and the image I 2 of
[0048] Step 2.3: Use the following formula to calculate the two images Recalculate Use gradient descent to update Net1 parameters, and use moving average (EMA) to update Net2 parameters;
[0049]
[0050] in, represents the mean function, h represents the feature encoding function based on conditional probability density, and Represents image I 1 The output value of Net1 and Net2, j represents The jth sample drawn from , k represents the number of samples drawn.
[0051] Step 2.4: Finally, the trained model is used to detect time series images to obtain a probability distribution diagram, and then a binary classification method is used to classify the change and unchanged categories. Generally, the Ostu or K-means clustering algorithm is used.
[0052] Step 3: Figure 3 As shown, the unchanged area detected in step 2 is analyzed pixel by pixel in time series to determine the radiation unchanged pixels and radiation deterministic pixels, which is divided into the following steps:
[0053] Step 3.1: Convert the image acquisition time into annual days as the time series independent variable d, and the pixel radiation value as the dependent variable V rc ;
[0054]
[0055] Among them, V rc represents the pixel value of the rth row and cth column of the time series image, d represents the annual accumulation day of the time series image, a 1 , a 2 , a 3 , a 4 , a 5 Represents the time series model coefficients.
[0056] Step 3.2: If Figure 3 As shown in the figure, using three times the curve regression error as the threshold, the pixels with more than three times the error of the fitting time series model curve are automatically removed iteratively. The radiation values of such pixels are unstable in time series analysis and do not have the characteristics of radiation invariance and radiation determination. Therefore, the pixels left after removal are suitable for constructing a radiation benchmark for subsequent radiation correction.
[0057] Taking Wuhan as an example, the radiation benchmark for Wuhan is established:
[0058] (1) Collect homologous image data of Wuhan area for three years or more, and ensure that at least one image can be obtained every quarter or month in the entire Wuhan city, so as to obtain time series images of Wuhan area for three years or more.
[0059] (2): Perform invariance detection on the paired continuous images of the time series images in (1), so as to detect which areas in Wuhan City have not changed for 3 years or more, namely the invariant areas.
[0060] (3): Since there are abnormal pixel values in the unchanged area in (2), step 3 uses time series analysis to eliminate abnormal pixel values.
[0061] (4): The pixels remaining after elimination are the unchanged pixels or radiation deterministic pixels, which serve as the radiation reference for radiation correction.
[0062] Step 4: Figure 4 As shown, the image to be corrected is firstly subjected to absolute radiometric calibration and atmospheric correction, then geometrically corrected with the radiometric reference pixels established by the method in steps 1-3, and finally radiometric correction is performed in combination with chi-square distribution and linear regression. Figure 4The thin black straight line represents the linear regression result after using chi-square distribution to screen radiation benchmark pixels. The gray straight line and thick black straight line represent the linear regression results before and after time series analysis, respectively. The thin black straight line indicates that the RMSE of radiation correction after time series analysis and chi-square distribution screening is significantly reduced, which can effectively improve radiation consistency. The specific steps are as follows:
[0063] Step 4.1: First, the image to be corrected is subjected to absolute radiometric calibration and atmospheric correction, and then geometrically corrected with the radiometric reference pixels established using the method in steps 1-3 to achieve sub-pixel accuracy of geographic coordinates, and then the geographic coordinates between the radiometric reference and the image to be corrected are matched;
[0064] Step 4.2: Combine the matched pixels with linear regression to construct chi-square distribution and remove some unstable pixels;
[0065]
[0066]
[0067] Among them, a j and b j represents the relative radiation correction coefficient of the j-band image, and represents the radiation value of the i-th pixel of the j-band of the image to be corrected and the j-band of the reference image, n represents the number of pixels involved in the calculation, T represents the chi-square distribution constructed using the root mean square error of linear regression, and χ belongs to the symbol of the chi-square distribution;
[0068] Step 4.3: Set the chi-square distribution probability threshold and select the chi-square distribution probability value For pixels greater than the threshold, linear regression is performed on the screened pixels to determine the radiation correction linear model. Finally, the radiation value of the image to be corrected is substituted into the linear model for calculation and radiation correction is performed.
[0069] y j =a j x j +b j
[0070] Among them, x j ,y j Indicates the corresponding pixel value of the j band of the image to be corrected and the reference image, a j and b j Represents the relative radiation correction coefficient of the j-band image.
[0071] It should be understood that parts not elaborated in detail in this specification belong to the prior art.
[0072] It should be understood that the above description of the preferred embodiment is relatively specific and cannot be regarded as limiting the scope of protection of the patent of the present invention. Under the guidance of the present invention, ordinary technicians in this field can also construct adaptive radiation sample libraries for different regions without departing from the scope of protection of the claims of the present invention, which all fall within the scope of protection of the present invention. The scope of protection requested by the present invention shall be based on the attached claims.
Claims
1. A method for establishing a radiation benchmark and correcting radiation for high-resolution remote sensing images, characterized in that: The following steps are involved: Step 1: determine the study area and obtain the homologous image sequence of the area in recent years as the image for constructing the radiation benchmark of the area; Step 2: Based on the deep learning-based invariant detection method, the invariant regions of the homologous time series images are selected; The specific implementation of step 2 is as follows; The time series images are paired I1 and I2 and input into the invariant detection model based on deep learning to calculate and Conduct training. The specific training process is as follows: Step 2.1, randomly crop the input image as training data, input it into Net1 and Net2 networks respectively, Net1 and Net2 are Swin-transformer, ResNet or VGG models, and then perform softmax calculation to obtain the probability distribution; first input image I1 into Net1 and Net2 networks for training, obtain the probability distribution P1-1 and P1-2 of the training results of image I1, and calculate the probability distribution of image I1 Step 2.2: Input image I2 into Net1 and Net2 networks for training, obtain the probability distributions P2-1 and P2-2 of the training results of image I2, and calculate the probability distribution of image I2. Step 2.3, use the following formula to calculate the loss of the two images contrast , and then calculate Use gradient descent to update Net1 parameters, and use sliding average to update Net2 parameters; in, represents the mean function, h represents the feature encoding function based on conditional probability density, and Respectively represent the output values of image I1 in Net1 and Net2, and j represents The jth sample drawn from , k represents the number of samples drawn; Step 2.4, finally, use the trained model to detect the time series images, obtain the probability distribution map, and then use the binary classification method to obtain the change and no change categories, where the binary classification method uses the Ostu or K-means clustering algorithm; Step 3, performing a time series analysis on the pixels in the unchanged area to determine the radiation unchanged pixels and the radiation deterministic pixels, which are used as radiation benchmarks for radiation correction; Step 4: First, preprocess the image to be corrected, then construct a chi-square distribution with the radiation benchmark determined in steps 1-3, perform statistical linear regression, remove abnormal pixels, and improve the relative radiation correction accuracy.
2. The method for establishing a radiation benchmark and correcting radiation for high-resolution remote sensing images according to claim 1 is characterized by: In step 1, the homologous image sequence of the study area for the past three years or more is used as the image for constructing the radiation benchmark of the area. Specifically, the Sentinel-2 image data is used as the homologous image sequence. The homologous image meets the following conditions: (1) To reduce the radiation differences caused by the systematic errors of the imaging equipment, images from the same satellite and the same sensor were selected; (2) The study area should contain image data from the past three years or more, and ensure that an image can be obtained every quarter or month.
3. The method for establishing a radiation benchmark and correcting radiation for high-resolution remote sensing images according to claim 1 is characterized by: The specific implementation of step 3 is as follows; Step 3.1: Convert the image acquisition time into annual days as the time series independent variable d, and the pixel radiation value as the dependent variable V rc ; Among them, V rc represents the pixel value of the rth row and cth column of the time series image, d represents the annual accumulation day of the time series image, a1, a2, a3, a4, a5 represent the time series model coefficients; In step 3.2, using three times the curve regression error as the threshold, the pixels with more than three times the error of the fitting time series model curve are automatically removed iteratively. The pixels left after removal are suitable for constructing the radiation benchmark for subsequent radiation correction.
4. The method for establishing a radiation benchmark and correcting radiation for high-resolution remote sensing images according to claim 1 is characterized by: The specific implementation of step 4 is as follows; Step 4.1, firstly, the image to be corrected is subjected to absolute radiometric calibration and atmospheric correction, and then geometrically corrected with the radiometric reference pixels established by steps 1-3 to achieve sub-pixel matching accuracy of geographic coordinates, and then the geographic coordinates between the radiometric reference and the image to be corrected are matched; Step 4.2, combine the matched pixels with linear regression to construct chi-square distribution and remove some unstable pixels; Among them, a j and b j represents the relative radiation correction coefficient of the j-band image, and represents the radiation value of the i-th pixel of the j-band of the image to be corrected and the j-band of the reference image, n represents the number of pixels involved in the calculation, T represents the chi-square distribution constructed using the root mean square error of linear regression, and χ belongs to the symbol of the chi-square distribution; In step 4.3, linear regression is performed on the pixels filtered by the chi-square distribution to determine the linear model of radiation correction. Finally, the radiation value of the image to be corrected is substituted into the linear model for calculation to perform radiation correction.