Hyperspectral remote sensing image relative radiation normalization method based on pseudo-invariant region pair

Abstract

The invention relates to a hyperspectral remote sensing image radiation normalization method, in particular to a hyperspectral remote sensing image relative radiation normalization method based on pseudo-invariant region pairs, and solves the technical problems that an existing pixel-based relative radiation normalization method is low in space matching precision and has spectral feature deviation. According to the method, a pseudo-invariant region is screened by adopting super-pixel segmentation, morphological processing and super-pixel guiding methods, then relative radiation normalization is performed on a to-be-corrected highlight remote sensing spectrum image through adaptive learning by adopting a self-correction model based on spatial spectrum constraint, and a quantitative method is adopted to correct the relative radiation of the to-be-corrected highlight remote sensing spectrum image. Not only can the spatial spectrum characteristics of the hyperspectral remote sensing image be fully utilized, but also the radiation intensity of the hyperspectral remote sensing image under different time phases can be adaptively corrected, and the consistency between time sequences and the quantitative inversion precision are improved, so that the accuracy and consistency of relative radiation normalization are improved, and the accuracy and the reliability of the relative radiation normalization are improved. The method is suitable for radiation normalization of hyperspectral remote sensing images with high spatial resolution and complex ground object scenes.

Description

Technical Field

[0001] The invention relates to a hyperspectral remote sensing image radiation normalization method, and in particular to a hyperspectral remote sensing image relative radiation normalization method based on pseudo-invariant region pairs. Background Art

[0002] Airborne hyperspectral remote sensing images have rich spectral information and high-resolution spatial information, and have great application potential in scenarios such as dynamic monitoring of land attributes, sewage detection, and quantitative inversion of land features. Relative radiation normalization is an important prerequisite for its quantitative application.

[0003] Traditional relative radiation normalization methods can be divided into two categories: pixel-based relative radiation normalization and distribution-based relative radiation normalization. At present, for the relative radiation normalization of the same hyperspectral load, researchers usually choose the pixel-based relative radiation normalization method to obtain better results. However, with the continuous improvement of the performance of hyperspectral detectors, the relative radiation normalization of high spatial resolution hyperspectral remote sensing images faces the following shortcomings:

[0004] (1) It is highly subjective. The pixel-based relative radiation normalization method often relies on artificially set parameters or thresholds during the processing process, which leads to a high degree of subjectivity in the normalization results, affecting their accuracy and consistency. This is especially true when processing high-spatial-resolution images. This subjectivity becomes more obvious due to the increase in image details.

[0005] (2) After extracting pseudo-invariant feature points, the least squares method or orthogonal regression method is usually used to calculate the mapping relationship between the hyperspectral remote sensing image to be corrected and the reference hyperspectral remote sensing image. For multi-phase hyperspectral remote sensing images after atmospheric correction, the main reason for the inconsistency of the spectrum between phases includes the incomplete stripping of the atmosphere, and is also affected by the noise of instruments and observations. Therefore, there is a nonlinear relationship between the two. When processing data with nonlinear relationships, if the least squares method or orthogonal regression method that assumes that there is a linear relationship between the data is used, the fitting results will be inaccurate, thereby affecting the accuracy of the subsequent normalization results;

[0006] (3) The band-by-band normalization method is usually adopted, which ignores the consistency of spectral shapes between bands, resulting in the deviation of the normalized spectral characteristics. Summary of the invention

[0007] The purpose of the present invention is to solve the technical problems of low spatial matching accuracy and spectral feature offset of the existing pixel-based relative radiation normalization method, and to provide a hyperspectral remote sensing image relative radiation normalization method based on pseudo-invariant region pairs.

[0008] To achieve the above object, the technical solution adopted by the present invention is:

[0009] A relative radiation normalization method for hyperspectral remote sensing images based on pseudo-invariant region pairs is characterized in that it includes the following steps:

[0010] Step 1: Use a spectrometer to obtain a reference hyperspectral remote sensing image DN_T1 and a hyperspectral remote sensing image to be corrected DN_T2 at different times, and then calculate the entrance pupil radiance of DN_T1 and DN_T2 respectively according to the calibration coefficient of the spectrometer;

[0011] Step 2, using the entrance pupil radiance of DN_T1 to perform atmospheric correction on DN_T1, using the entrance pupil radiance of DN_T2 to perform atmospheric correction on DN_T2, and then performing image registration on the two to achieve feature space alignment, and obtain feature space aligned DN_T1 and DN_T2;

[0012] Step 3, preprocessing the DN_T1 and DN_T2 aligned in feature space respectively, removing the regions that change significantly over time, and obtaining enhanced DN_T1 and DN_T2;

[0013] Step 4: Use a characteristic band selection network based on the spectral response function to perform dimensionality reduction processing on the enhanced DN_T1 and DN_T2 respectively, extract their spectral features, and obtain the reduced-dimensional DN_T1 and DN_T2;

[0014] Step 5: Using an iterative weighted multivariate change detection transformation algorithm, obtain pseudo-invariant feature point groups of DN_T1 and DN_T2 in enhanced DN_T1 and DN_T2 respectively;

[0015] Step 6: Use superpixel segmentation to cluster the features of DN_T1 and DN_T2 after dimension reduction, perform morphological processing on the pseudo-invariant feature point groups of DN_T1 and DN_T2, and then select the pseudo-invariant regions of DN_T1 and DN_T2 according to the feature clustering results and the morphological processing results;

[0016] Step 7, pair the pseudo-invariant regions of DN_T1 and DN_T2, and then input the paired pseudo-invariant regions into the self-correction initial model based on the spatial spectrum constraint, the self-correction initial model normalizes the pseudo-invariant region of DN_T2, and then adjusts the parameters of the self-correction initial model according to the loss function value between the normalization result and the pseudo-invariant region of DN_T1, to obtain the self-correction model based on the spatial spectrum constraint;

[0017] Step 8: Input the feature space aligned DN_T2 into the self-correction model based on spatial spectrum constraints to obtain the normalized hyperspectral remote sensing image and complete the relative radiation normalization of the hyperspectral remote sensing image.

[0018] Furthermore, step 6 is specifically as follows:

[0019] 6.1. Use the superpixel segmentation algorithm to divide the ground feature information in DN_T1 and DN_T2 after dimension reduction into superpixels, and obtain the superpixel region sets of DN_T1 and DN_T2;

[0020] 6.2. Set the spectral angle distance threshold and the root mean square error threshold, calculate the spectral angle distance and the root mean square error of two adjacent superpixel regions in the DN_T1 and DN_T2 superpixel region sets respectively, if the spectral angle distance of two adjacent superpixel regions is less than the spectral angle distance threshold and the root mean square error is less than the root mean square error threshold, then merge the two adjacent superpixel regions to obtain the optimized region sets of DN_T1 and DN_T2 respectively;

[0021] 6.3. Use morphological methods to process the pseudo-invariant feature point groups of DN_T1 and DN_T2 to obtain the initial set of pseudo-invariant regions of DN_T1 and DN_T2;

[0022] 6.4. Take the intersection of the optimized region set and the initial set of pseudo-invariant regions of DN_T1, and the optimized region set and the initial set of pseudo-invariant regions of DN_T2, and then calculate the maximum inscribed rectangle of the intersection to obtain the pseudo-invariant regions of DN_T1 and DN_T2.

[0023] Furthermore, step 4 is specifically as follows:

[0024] 4.1. Construct a feature band selection network based on the spectral response function and set a loss function threshold of the feature band selection network; the feature band network includes an encoding network and a decoding network connected in sequence, and the decoding network is based on the spectral Transforme technology;

[0025] 4.2. The enhanced DN_T1 or DN_T2 is input into the encoding network, the encoding network simulates the spectral response curve, selects the spectral features sensitive to the spectral response in the enhanced DN_T1 or DN_T2, thereby reducing the dimension of the enhanced DN_T1 or DN_T2 to obtain the low-dimensional features of DN_T1 or DN_T2;

[0026] 4.3. The decoding network reconstructs the low-dimensional features of DN_T1 or DN_T2 to obtain the feature cube of DN_T1 or DN_T2;

[0027] 4.4. Calculate the loss function value of the feature band selection network according to the feature cube of DN_T1 or DN_T2, and adjust the parameters of the encoding network and the decoding network according to the loss function value, and then return to step 4.2 to iterate until the loss function value is less than the loss function threshold, and obtain the trained feature band selection network;

[0028] 4.5. Input the enhanced DN_T1 and DN_T2 into the trained feature band selection network to obtain the reduced-dimensional DN_T1 and DN_T2.

[0029] Furthermore, in step 4.1, the encoding network includes a 1×1 3d convolution unit, a LeakyReLU activation unit, and a residual link unit connected in sequence.

[0030] Furthermore, in step 4.1, the loss function of the feature band selection network is the root mean square error between the output of the decoding network and the input of the encoding network.

[0031] Further, in step 7, the self-correction model based on spatial spectrum constraint includes an input layer, a feature extraction layer, a spectrum constraint layer, a radiation difference learning layer and a radiation correction output layer connected in sequence;

[0032] The feature extraction layer includes a 3D convolution residual block for extracting spatial features and spectral features of pseudo-invariant regions, and the spectral constraint layer is based on a spectral attention mechanism.

[0033] Furthermore, in step 7, the loss function L of the self-correction model based on the space spectrum constraint is * for:

[0034] L * =α*D KL (T2',T1)+β*SAM(mean(T2'),mean(T1))+γ*SSIM(T1,T1')

[0035] Among them, D KL (T2′, T1) represents the KL divergence distribution between the pseudo-invariant region of DN_T2 and the pseudo-invariant region of DN_T1 after relative radiation normalization, T2′ represents the pseudo-invariant region of DN_T2 after relative radiation normalization, and T1 represents the pseudo-invariant region of DN_T1;

[0036] SAM(mean(T2′),mean(T1)) represents the spectral angular distance between the spectral mean of the pseudo-invariant region of DN_T2 calculated in the channel dimension after relative radiation normalization and the spectral mean of DN_T1, mean(T2′) represents the spectral mean of the pseudo-invariant region of DN_T2 calculated in the channel dimension after relative radiation normalization, and mean(T1) represents the spectral mean of DN_T1;

[0037] SSIM(T1′,T1) represents the structural similarity between DN_T1 and DN_T1 after relative radiation normalization, and T1′ represents the pseudo-invariant region of DN_T1 after relative radiation normalization;

[0038] α, β, and γ are respectively KL Weight parameters of SSIM(T2′,T1), SAM(mean(T2′),mean(T1)), and SSIM(T1′,T1).

[0039] Furthermore, in step 3, the areas that change significantly over time include vegetation-covered areas, water areas, and shadow areas, and step 3 is specifically as follows:

[0040] 3.1. Extract characteristic spectrum segments of vegetation coverage area and water area in the database respectively;

[0041] 3.2. Calculate the normalized vegetation index and normalized difference water index based on the characteristic spectrum of the vegetation coverage area and the water area respectively;

[0042] 3.3. Divide the DN_T1 and DN_T2 with feature space alignment into regions, and set the vegetation cover threshold, water body threshold and shadow threshold;

[0043] 3.4. Calculate the ratio of the difference between the reflectance of the near-infrared band and the red band of each area and the sum of the two according to the normalized vegetation index. If the ratio is greater than the vegetation coverage area threshold, the corresponding area will be eliminated;

[0044] The reflectance difference between the shortwave infrared band and the green light band of each area is calculated according to the normalized difference water index. If the reflectance difference is greater than the water body threshold, the corresponding area will be eliminated.

[0045] The shadow index of each area is calculated, and if the shadow index is greater than the shadow threshold, the corresponding area is removed. Further, in step 1, the entrance pupil radiance is calculated by the following formula:

[0046] L=a*DN+b

[0047] Among them, L is the entrance pupil radiance, DN is the energy intensity of the hyperspectral remote sensing image, and a and b are the calibration coefficients of the spectrometer.

[0048] Furthermore, in step 2, the image registration method is: combining POS information and image SURF feature points for image registration.

[0049] Compared with the prior art, the present invention has the following beneficial technical effects:

[0050] 1. The method for relative radiation normalization of hyperspectral remote sensing images based on pseudo-invariant region pairs provided by the present invention is suitable for hyperspectral remote sensing images with high spatial resolution and complex ground scenes. It uses superpixel segmentation and morphological processing to screen pseudo-invariant regions, and uses a self-correction model based on spatial-spectral constraints to perform relative radiation normalization on the hyperspectral remote sensing images to be corrected through adaptive learning. The quantitative method can not only make full use of the spatial-spectral characteristics of the hyperspectral remote sensing images, but also adaptively correct the radiation intensity of the hyperspectral remote sensing images in different phases, which can improve the consistency between phases and the quantitative inversion accuracy, solve the problems of low spatial matching accuracy and spectral feature offset, and improve the accuracy and consistency of relative radiation normalization;

[0051] 2. In the relative radiation normalization method of hyperspectral remote sensing images based on pseudo-invariant region pairs provided by the present invention, morphological processing and superpixel segmentation methods are used to guide the screening of pseudo-invariant regions through superpixels, which can retain rich spatial spectrum features and extract characteristic relationships of spatial neighborhoods, thereby reducing the interference of neighboring effects on the screening of pseudo-invariant regions;

[0052] 3. In the relative radiation normalization method of hyperspectral remote sensing images based on pseudo-invariant region pairs provided by the present invention, the encoding network simulates the spectral response curve and selects spectral features that are sensitive to the spectral response, which can ensure the maximization of information after dimensionality reduction; the decoding network is based on spectral Transforme technology, which can better understand the complex relationship between spectral features, thereby generating more accurate reconstructed data;

[0053] 4. In the relative radiation normalization method of hyperspectral remote sensing images based on pseudo-invariant region pairs provided by the present invention, the self-correction model based on spatial spectrum constraints includes a spectral constraint layer, and introduces a spectral attention mechanism to focus on the regions or bands that have a greater impact on radiation normalization, thereby improving the accuracy of relative radiation normalization;

[0054] 5. In the relative radiation normalization method of hyperspectral remote sensing images based on pseudo-invariant region pairs provided by the present invention, the loss function of the self-correction model based on space-spectrum constraints includes KL divergence and spectral angular distance, which can not only maintain the consistency of space and spectrum, but also avoid spatial misalignment leading to spatial transformation of the hyperspectral remote sensing image to be corrected based on the guidance of the reference hyperspectral remote sensing image;

[0055] 6. In the relative radiation normalization method of hyperspectral remote sensing images based on pseudo-invariant region pairs provided by the present invention, the reference hyperspectral remote sensing images and the hyperspectral remote sensing images to be corrected are screened by using the normalized vegetation index, the normalized difference water index and the shadow index. This can eliminate interference information in the process of extracting pseudo-invariant feature point groups and ensure the accuracy of subsequent relative radiation normalization. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 is a flow chart of a method according to an embodiment of the present invention;

[0057] Figure 2 The reference hyperspectral remote sensing image obtained in step 1 in the embodiment of the present invention;

[0058] Figure 3 The hyperspectral remote sensing image to be corrected obtained in step 1 in the embodiment of the present invention;

[0059] Figure 4 The normalized hyperspectral remote sensing image obtained in step 8 of the embodiment of the present invention;

[0060] Figure 5 It is a spectrum curve diagram of the reference hyperspectral remote sensing image, the hyperspectral remote sensing image to be corrected, and the normalized hyperspectral remote sensing image in the embodiment of the present invention. DETAILED DESCRIPTION

[0061] The following is a further detailed description of a method for normalizing relative radiation of hyperspectral remote sensing images based on pseudo-invariant region pairs proposed by the present invention in conjunction with the accompanying drawings and specific embodiments. It should be understood by those skilled in the art that these embodiments are only used to explain the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0062] A relative radiometric normalization method for hyperspectral remote sensing images based on pseudo-invariant region pairs, such as Figure 1 As shown, the following steps are included:

[0063] Step 1: Use a spectrometer to obtain Figure 2 The reference hyperspectral remote sensing image DN_T1 and Figure 3 The hyperspectral remote sensing image DN_T2 to be corrected is shown, and then the entrance pupil radiance of DN_T1 and DN_T2 is calculated respectively according to the calibration coefficient of the spectrometer. The entrance pupil radiance is calculated by the following formula:

[0064] L=a*DN+b

[0065] Among them, L is the entrance pupil radiance, DN is the energy intensity of the hyperspectral remote sensing image, and a and b are the calibration coefficients of the spectrometer.

[0066] Step 2: Use the entrance pupil radiance of DN_T1 to perform atmospheric correction on DN_T1, and use the entrance pupil radiance of DN_T2 to perform atmospheric correction on DN_T2. Then, combine the POS information and the image SURF feature points to perform image registration on the two to achieve feature space alignment, and obtain DN_T1 and DN_T2 aligned in feature space.

[0067] Step 3: Preprocess the feature space aligned DN_T1 and DN_T2 respectively, remove the areas that change significantly over time, and obtain enhanced DN_T1 and DN_T2.

[0068] Among them, the areas that change significantly over time include vegetation-covered areas, water areas, and shadow areas. Step 3 is specifically as follows:

[0069] 3.1. Extract characteristic spectrum segments of vegetation coverage area and water area in the database respectively;

[0070] 3.2. Calculate the normalized vegetation index and normalized difference water index based on the characteristic spectrum of the vegetation coverage area and the water area respectively;

[0071] 3.3. Divide the DN_T1 and DN_T2 with feature space alignment into regions, and set the vegetation cover threshold, water body threshold and shadow threshold;

[0072] 3.4. Calculate the ratio of the difference between the reflectance of the near-infrared band and the red band of each area and the sum of the two according to the normalized vegetation index. If the ratio is greater than the vegetation coverage area threshold, the corresponding area will be eliminated;

[0073] The reflectance difference between the shortwave infrared band and the green light band of each area is calculated according to the normalized difference water index. If the reflectance difference is greater than the water body threshold, the corresponding area will be eliminated.

[0074] The shadow index of each area is calculated. If the shadow index is greater than the shadow threshold, the corresponding area is removed.

[0075] Step 4: Use the feature selection band network based on the spectral response function to reduce the dimension of the enhanced DN_T1 and DN_T2, extract their spectral features, and obtain the reduced-dimensional DN_T1 and DN_T2. Specifically:

[0076] 4.1. Construct a feature band selection network based on the spectral response function and set a loss function threshold of the feature band selection network; the feature band network includes an encoding network and a decoding network connected in sequence, and the decoding network is based on the spectral Transforme technology;

[0077] 4.2. The enhanced DN_T1 or DN_T2 is input into the encoding network, the encoding network simulates the spectral response curve, selects the spectral features sensitive to the spectral response in the enhanced DN_T1 or DN_T2, thereby reducing the dimension of the enhanced DN_T1 or DN_T2 to obtain the low-dimensional features of DN_T1 or DN_T2;

[0078] 4.3. The decoding network reconstructs the low-dimensional features of DN_T1 or DN_T2 to obtain the feature cube of DN_T1 or DN_T2;

[0079] 4.4. Calculate the loss function value of the feature band selection network according to the feature cube of DN_T1 or DN_T2, and adjust the parameters of the encoding network and the decoding network according to the loss function value, and then return to step 4.2 to iterate until the loss function value is less than the loss function threshold, and obtain the trained feature band selection network;

[0080] 4.5. Input the enhanced DN_T1 and DN_T2 into the trained feature band selection network to obtain the reduced-dimensional DN_T1 and DN_T2.

[0081] In this step, the encoding network is composed of a stack of residual modules, which include 1×1 3d convolution units, LeakyReLU activation units, and residual link units connected in sequence. The encoding network compresses the input data into a low-dimensional feature representation. The decoding network is based on spectral Transforme technology, which can capture long-distance dependencies in the data and help the decoding network better understand the complex relationship between spectral features, thereby generating more accurate reconstructed data. The loss function of the feature band selection network uses the root mean square error between the output of the decoding network and the input of the encoding network to guide the feature band selection network to extract low-dimensional features.

[0082] Step 5: Use the iterative weighted multivariate change detection transformation algorithm to obtain the pseudo-invariant feature point groups of DN_T1 and DN_T2 in the enhanced DN_T1 and DN_T2 respectively.

[0083] Step 6: Use superpixel segmentation to cluster the features of DN_T1 and DN_T2 after dimension reduction, and perform morphological processing on the pseudo-invariant feature point groups of DN_T1 and DN_T2, and then select the pseudo-invariant areas of DN_T1 and DN_T2 according to the feature clustering results and morphological processing results. Specifically:

[0084] 6.1. Use the superpixel segmentation algorithm to divide the ground feature information in DN_T1 and DN_T2 after dimension reduction into superpixels, and obtain the superpixel region sets of DN_T1 and DN_T2;

[0085] 6.2. Set the spectral angle distance threshold and the root mean square error threshold, calculate the spectral angle distance and the root mean square error of two adjacent superpixel regions in the DN_T1 and DN_T2 superpixel region sets respectively, if the spectral angle distance of two adjacent superpixel regions is less than the spectral angle distance threshold and the root mean square error is less than the root mean square error threshold, then merge the two adjacent superpixel regions to obtain the optimized region sets of DN_T1 and DN_T2 respectively;

[0086] 6.3. Use morphological methods to process the pseudo-invariant feature point groups of DN_T1 and DN_T2 to obtain the initial set of pseudo-invariant regions of DN_T1 and DN_T2;

[0087] 6.4. Take the intersection of the optimized region set and the initial set of pseudo-invariant regions of DN_T1, and the optimized region set and the initial set of pseudo-invariant regions of DN_T2, and then calculate the maximum inscribed rectangle of the intersection to obtain the pseudo-invariant regions of DN_T1 and DN_T2.

[0088] Step 7, pair the pseudo-invariant regions of DN_T1 and DN_T2, and then input the paired pseudo-invariant regions into the self-correction initial model based on spatial spectrum constraints. The self-correction initial model normalizes the pseudo-invariant region of DN_T2, and then adjusts the parameters of the self-correction initial model according to the loss function value between the normalization result and the pseudo-invariant region of DN_T1 to obtain the self-correction model based on spatial spectrum constraints.

[0089] Among them, the self-correction model based on spatial-spectral constraints includes an input layer, a feature extraction layer, a spectral constraint layer, a radiation difference learning layer and a radiation correction output layer connected in sequence. The feature extraction layer includes a 3D convolution residual block for extracting the spatial features and spectral features of the pseudo-invariant region. The input layer is used to input the pseudo-invariant regions of DN_T1 and DN_T2; the feature extraction layer is used to extract the spatial features and spectral features of the pseudo-invariant region; the spectral constraint layer introduces a spectral attention mechanism to help the autoencoder model focus on the regions or bands that have a greater impact on radiation normalization; the radiation difference learning layer is used to capture global radiation changes; the radiation correction output layer is used to output the radiation normalized image, so that the normalized hyperspectral remote sensing image to be corrected is consistent with the reference hyperspectral remote sensing image in terms of intensity, while retaining the spatial and spectral dimensions of the original image.

[0090] The loss function L of the self-correcting model * for:

[0091] L * =α*D KL (T2',T1)+β*SAM(mean(T2'),mean(T1))+γ*SSIM(T1,T1')

[0092] Among them, D KL (T2′, T1) represents the KL divergence distribution between the pseudo-invariant region of DN_T2 and the pseudo-invariant region of DN_T1 after relative radiation normalization, T2′ represents the pseudo-invariant region of DN_T2 after relative radiation normalization, and T1 represents the pseudo-invariant region of DN_T1;

[0093] SAM(mean(T2′),mean(T1)) represents the spectral angular distance between the spectral mean of the pseudo-invariant region of DN_T2 calculated in the channel dimension after relative radiation normalization and the spectral mean of DN_T1, mean(T2′) represents the spectral mean of the pseudo-invariant region of DN_T2 calculated in the channel dimension after relative radiation normalization, and mean(T1) represents the spectral mean of DN_T1;

[0094] SSIM(T1′,T1) represents the structural similarity between DN_T1 and DN_T1 after relative radiation normalization, and T1′ represents the pseudo-invariant region of DN_T1 after relative radiation normalization;

[0095] α, β, and γ are respectively KL The weight parameters of SAM(mean(T2′),mean(T1)) and SSIM(T1′,T1) are as follows: in this embodiment, the initial values ​​of α, β and γ are 0.6, 0.2 and 0.2 respectively.

[0096] In the loss function of the self-correction model, the KL divergence distribution is introduced to constrain the normalized DN_T2 intensity distribution to be consistent with the intensity distribution of DN_T1; the spectral angular distance is introduced to avoid the change of the spectral characteristics of DN_T2 during the correction process; the structural similarity is introduced to ensure that the spatial texture of the normalized DN_T2 and the original image remains unchanged.

[0097] Step 8: Input the feature space aligned DN_T2 into the self-correction model based on the space spectrum constraint to obtain Figure 4 The normalized hyperspectral remote sensing image shown completes the relative radiation normalization of the hyperspectral remote sensing image.

[0098] like Figure 2-Figure 4 As shown in Figure 2, compared with the hyperspectral remote sensing image DN_T2 to be corrected, the radiation intensity of the corrected image is closer to the radiation intensity of the reference hyperspectral remote sensing image DN_T1. Figure 5 As shown, the spectrum of the unchanged area is selected for comparison. It can be seen that compared with the hyperspectral remote sensing image DN_T2 to be corrected, the spectrum of the normalized hyperspectral remote sensing image is closer to the spectrum of the reference hyperspectral remote sensing image DN_T1. It can be seen that the relative radiation normalization method provided in this embodiment has high accuracy and consistency.

[0099] The relative radiation normalization method for hyperspectral remote sensing images based on pseudo-invariant region pairs provided in this embodiment makes full use of the spatial-spectral characteristics of high-spatial-resolution hyperspectral remote sensing images, and at the same time constructs a spatial-spectral constraint loss function for relative radiation normalization under incomplete spatial alignment. It adopts a quantitative method without complicated operations, and can adaptively correct the radiation intensity of hyperspectral remote sensing images in different phases, thereby improving the consistency between time series and the accuracy of relative radiation normalization.

Claims

1. A method for relative radiometric normalization of hyperspectral remote sensing images based on pseudo-invariant region pairs, characterized in that: The following steps are involved: Step 1: Use a spectrometer to obtain a reference hyperspectral remote sensing image DN_T1 and a hyperspectral remote sensing image to be corrected DN_T2 at different times, and then calculate the entrance pupil radiance of DN_T1 and DN_T2 respectively according to the calibration coefficient of the spectrometer; Step 2, using the entrance pupil radiance of DN_T1 to perform atmospheric correction on DN_T1, using the entrance pupil radiance of DN_T2 to perform atmospheric correction on DN_T2, and then performing image registration on the two to achieve feature space alignment, and obtain feature space aligned DN_T1 and DN_T2; Step 3, preprocessing the DN_T1 and DN_T2 aligned in feature space respectively, removing the regions that change significantly over time, and obtaining enhanced DN_T1 and DN_T2; Step 4: Use a characteristic band selection network based on the spectral response function to perform dimensionality reduction processing on the enhanced DN_T1 and DN_T2 respectively, extract their spectral features, and obtain the reduced-dimensional DN_T1 and DN_T2; Step 5: Using an iterative weighted multivariate change detection transformation algorithm, obtain pseudo-invariant feature point groups of DN_T1 and DN_T2 in enhanced DN_T1 and DN_T2 respectively; Step 6: Use superpixel segmentation to cluster the features of DN_T1 and DN_T2 after dimension reduction, perform morphological processing on the pseudo-invariant feature point groups of DN_T1 and DN_T2, and then select the pseudo-invariant regions of DN_T1 and DN_T2 according to the feature clustering results and the morphological processing results; Step 7, pair the pseudo-invariant regions of DN_T1 and DN_T2, and then input the paired pseudo-invariant regions into the self-correction initial model based on the spatial spectrum constraint, the self-correction initial model normalizes the pseudo-invariant region of DN_T2, and then adjusts the parameters of the self-correction initial model according to the loss function value between the normalization result and the pseudo-invariant region of DN_T1, to obtain the self-correction model based on the spatial spectrum constraint; Step 8: Input the feature space aligned DN_T2 into the self-correction model based on spatial spectrum constraints to obtain the normalized hyperspectral remote sensing image and complete the relative radiation normalization of the hyperspectral remote sensing image.

2. The method for relative radiometric normalization of hyperspectral remote sensing images based on pseudo-invariant region pairs according to claim 1, characterized in that: Step 6 is as follows: 6.

1. Use the superpixel segmentation algorithm to divide the ground feature information in DN_T1 and DN_T2 after dimension reduction into superpixels, and obtain the superpixel region sets of DN_T1 and DN_T2; 6.

2. Set the spectral angle distance threshold and the root mean square error threshold, calculate the spectral angle distance and the root mean square error of two adjacent superpixel regions in the DN_T1 and DN_T2 superpixel region sets respectively, if the spectral angle distance of two adjacent superpixel regions is less than the spectral angle distance threshold and the root mean square error is less than the root mean square error threshold, then merge the two adjacent superpixel regions to obtain the optimized region sets of DN_T1 and DN_T2 respectively; 6.

3. Use morphological methods to process the pseudo-invariant feature point groups of DN_T1 and DN_T2 to obtain the initial set of pseudo-invariant regions of DN_T1 and DN_T2; 6.

4. Take the intersection of the optimized region set and the initial set of pseudo-invariant regions of DN_T1, and the optimized region set and the initial set of pseudo-invariant regions of DN_T2, and then calculate the maximum inscribed rectangle of the intersection to obtain the pseudo-invariant regions of DN_T1 and DN_T2.

3. The method for relative radiometric normalization of hyperspectral remote sensing images based on pseudo-invariant region pairs according to claim 2, characterized in that: Step 4 is as follows: 4.

1. Construct a feature band selection network based on the spectral response function and set a loss function threshold of the feature band selection network; the feature band network includes an encoding network and a decoding network connected in sequence, and the decoding network is based on the spectral Transforme technology; 4.

2. The enhanced DN_T1 or DN_T2 is input into the encoding network, the encoding network simulates the spectral response curve, selects the spectral features sensitive to the spectral response in the enhanced DN_T1 or DN_T2, thereby reducing the dimension of the enhanced DN_T1 or DN_T2 to obtain the low-dimensional features of DN_T1 or DN_T2; 4.

3. The decoding network reconstructs the low-dimensional features of DN_T1 or DN_T2 to obtain the feature cube of DN_T1 or DN_T2; 4.

4. Calculate the loss function value of the feature band selection network according to the feature cube of DN_T1 or DN_T2, and adjust the parameters of the encoding network and the decoding network according to the loss function value, and then return to step 4.2 to iterate until the loss function value is less than the loss function threshold, and obtain the trained feature band selection network; 4.

5. Input the enhanced DN_T1 and DN_T2 into the trained feature band selection network to obtain the reduced-dimensional DN_T1 and DN_T2.

4. The method for relative radiometric normalization of hyperspectral remote sensing images based on pseudo-invariant region pairs according to claim 3, characterized in that: In step 4.1, the encoding network includes a 1×1 3d convolution unit, a LeakyReLU activation unit, and a residual link unit connected in sequence.

5. The method for relative radiometric normalization of hyperspectral remote sensing images based on pseudo-invariant region pairs according to claim 4, characterized in that: In step 4.1, the loss function of the feature band selection network is the root mean square error between the output of the decoding network and the input of the encoding network.

6. The method for relative radiometric normalization of hyperspectral remote sensing images based on pseudo-invariant region pairs according to any one of claims 1 to 5, characterized in that: In step 7, the self-correction model based on spatial spectrum constraint includes an input layer, a feature extraction layer, a spectrum constraint layer, a radiation difference learning layer and a radiation correction output layer connected in sequence; The feature extraction layer includes a 3D convolution residual block for extracting spatial features and spectral features of pseudo-invariant regions, and the spectral constraint layer is based on a spectral attention mechanism.

7. The method for relative radiometric normalization of hyperspectral remote sensing images based on pseudo-invariant region pairs according to claim 6, characterized in that: In step 7, the loss function L of the self-correction model based on the space spectrum constraint is * for: 50 * =α*D KL (T2',T1)+β*SAM(mean(T2'),mean(T1)(+γ*SSIM(T1,T1') Among them, D KL (T2′, T1) represents the KL divergence distribution between the pseudo-invariant region of DN_T2 and the pseudo-invariant region of DN_T1 after relative radiation normalization, T2′ represents the pseudo-invariant region of DN_T2 after relative radiation normalization, and T1 represents the pseudo-invariant region of DN_T1; SAM(mean(T2′),mean(T1)) represents the spectral angular distance between the spectral mean of the pseudo-invariant region of DN_T2 calculated in the channel dimension after relative radiation normalization and the spectral mean of DN_T1, mean(T2′) represents the spectral mean of the pseudo-invariant region of DN_T2 calculated in the channel dimension after relative radiation normalization, and mean(T1) represents the spectral mean of DN_T1; SSIM(T1′,T1) represents the structural similarity between DN_T1 and DN_T1 after relative radiation normalization, and T1′ represents the pseudo-invariant region of DN_T1 after relative radiation normalization; α, β, and γ are respectively KL Weight parameters of SSIM(T2′,T1), SAM(mean(T2′),mean(T1)), and SSIM(T1′,T1).

8. The method for relative radiometric normalization of hyperspectral remote sensing images based on pseudo-invariant region pairs according to claim 7, characterized in that: In step 3, the areas that change significantly over time include vegetation-covered areas, water areas, and shadow areas. Step 3 is specifically as follows: 3.

1. Extract characteristic spectrum segments of vegetation coverage area and water area in the database respectively; 3.

2. Calculate the normalized vegetation index and normalized difference water index based on the characteristic spectrum of the vegetation coverage area and the water area respectively; 3.

3. Divide the DN_T1 and DN_T2 with feature space alignment into regions, and set the vegetation cover threshold, water body threshold and shadow threshold; 3.

4. Calculate the ratio of the difference between the reflectance of the near-infrared band and the red band of each area and the sum of the two according to the normalized vegetation index. If the ratio is greater than the vegetation coverage area threshold, the corresponding area will be eliminated; The reflectance difference between the shortwave infrared band and the green light band of each area is calculated according to the normalized difference water index. If the reflectance difference is greater than the water body threshold, the corresponding area will be eliminated. The shadow index of each area is calculated. If the shadow index is greater than the shadow threshold, the corresponding area is removed.

9. The method for relative radiometric normalization of hyperspectral remote sensing images based on pseudo-invariant region pairs according to claim 8, characterized in that: In step 1, the entrance pupil radiance is calculated by the following formula: L=a*DN+b Among them, L is the entrance pupil radiance, DN is the energy intensity of the hyperspectral remote sensing image, and a and b are the calibration coefficients of the spectrometer.

10. The method for relative radiometric normalization of hyperspectral remote sensing images based on pseudo-invariant region pairs according to claim 9, characterized in that: In step 2, the image registration method is: combining POS information and image SURF feature points for image registration.

Images