Remote sensing radiometric calibration target range extraction method fusing signal domain features

Abstract

The invention provides a remote sensing radiometric calibration target range extraction method fusing signal domain features, and belongs to the technical field of remote sensing science. The problem that the target range cannot be automatically screened in a large range in the existing remote sensing image calibration target range is solved; comprising the following steps: carrying out spatial-spectral distribution feature extraction on a multi-source time sequence remote sensing image, specifically, selecting and analyzing a feature wave band; analyzing spatial uniformity; analyzing time consistency; analyzing geographical distribution characteristics; a Mann-Kendall algorithm is adopted to carry out preprocessing before textural feature extraction on a multi-source time sequence remote sensing image, then textural features of the preprocessed multi-source time sequence remote sensing image are extracted based on two-dimensional empirical mode decomposition, and the textural features and spatial-spectral distribution features are fused to obtain basic feature information of the multi-source time sequence remote sensing image; performing unsupervised classification learning on the multi-source time sequence remote sensing image based on the fused basic feature information to obtain a target range distribution diagram; the method is applied to remote sensing radiometric calibration target range selection.

Description

Technical Field

[0001] The invention provides a remote sensing radiation calibration target field extraction method integrating signal domain features, belonging to the field of remote sensing science and technology. Background Art

[0002] An important prerequisite for the application of remote sensing images is the stability of their quality. Radiation calibration is the process of establishing a quantitative relationship between satellite sensor observations (count values, voltage values, etc.) and radiation physical quantities (radiance, reflectivity, etc.). It is a key link in satellite remote sensing technology. The selection of radiation calibration range is the basic work of radiation calibration. Good and accurate calibration results depend on the selection of a stable radiation calibration range with good ground and atmospheric properties.

[0003] The traditional method of selecting remote sensing radiation calibration sites mainly relies on manual evaluation of spatial uniformity and temporal consistency to select satisfactory areas, but there are still the following problems:

[0004] (1) Currently, the amount of data that needs to be recalibrated is still huge, and the requirements for satellite accuracy have also been further improved. The number of existing target ranges selected based on natural environments such as deserts and Gobi is slightly insufficient, and an intelligent selection method that can take into account a wide range of factors is needed.

[0005] (2) At present, in the evaluation of spatial uniformity of shooting range site characteristics, the use of coefficient of variation evaluation will lead to overly fragmented results and cannot reflect the overall characteristics of a large area. The use of Getis statistics cannot count the variance within the target window and cannot well reflect the spatial uniformity of the target area. A method is needed to comprehensively consider the selection of the shooting range.

[0006] As the number of satellite launches continues to increase, the quantity and quality of remote sensing images are also constantly improving. There is an urgent need for a method that can automatically screen target ranges over a large area and consider the characteristics of the ground and atmosphere more comprehensively. Summary of the invention

[0007] In order to solve the above technical problems, a remote sensing radiation calibration target range extraction method integrating signal domain features is provided, comprising the following steps:

[0008] A: Extract spatial-spectral distribution features of multi-source time-series remote sensing images, including:

[0009] a1) Characteristic band selection and analysis;

[0010] a2) Spatial uniformity analysis;

[0011] a3) Time consistency analysis;

[0012] a4) Analysis of geographical distribution characteristics;

[0013] B: The Mann-Kendall algorithm is used to preprocess the multi-source time series remote sensing images before texture feature extraction. Then, the texture features of the preprocessed multi-source time series remote sensing images are extracted based on two-dimensional empirical mode decomposition, and the texture features are fused with the spatial spectrum distribution features to obtain the basic feature information of the multi-source time series remote sensing images.

[0014] C: Based on the fused basic feature information, unsupervised classification learning is performed on multi-source time series remote sensing images to obtain the target range distribution map.

[0015] In step a1), the characteristic bands are selected by assigning weights to the original bands of the image, wherein the weights are calculated by weighted superposition of the information content of the image bands and the noise estimation value of the image.

[0016] The information content of the image band is obtained through grayscale statistics;

[0017] The noise estimation value of an image is the ratio of the noise value of the entire image to the grayscale mean.

[0018] The band weight is calculated as follows:

[0019]

[0020] In the formula: BI i is the weight of band i, n is the band index, H i is the information content of the i-band image, H min is the minimum information content of the image, H max is the maximum amount of information in the image, N min is the minimum noise estimate of the image, N max is the maximum noise estimate of the image, and the sum of all band weights is 1;

[0021] The blue band is seriously polluted by the atmosphere, so its weight is 0. For other bands, the weights are calculated according to the above formula.

[0022] In step a2), Getis statistic or CV index is used to judge the spatial uniformity of the multi-source time series remote sensing images preprocessed by the Mann-Kendall algorithm.

[0023] In step a3), the temporal consistency of the images is determined by statistically analyzing the mean and standard deviation of the grayscale values ​​of the time-series images.

[0024] The geographical distribution characteristics analysis in step a4) further includes: normalized vegetation index analysis and normalized building index analysis.

[0025] In step B, the texture features of the preprocessed multi-source time series remote sensing images are extracted based on two-dimensional empirical mode decomposition. The specific implementation steps are as follows:

[0026] The intrinsic mode function (IMF) of remote sensing images is obtained by two-dimensional empirical mode decomposition. The three IMF components except the first IMF are extracted. If the components are insufficient, the last IMF component is copied to supplement them. The overall weight of the three IMF components is 3.

[0027] The specific implementation steps of step C are as follows:

[0028] The extraction of the target range based on unsupervised learning classification requires the integration of the spatial uniformity, temporal consistency, geographical distribution characteristics and texture characteristics extracted previously, superimposing the original image and the extracted features, and completing the comprehensive consideration of the features through the constructed weights to obtain the characteristic remote sensing image. The characteristic remote sensing images are clustered using the K-means clustering method, and the clustering results are analyzed after the morphological operation to obtain the final remote sensing radiation calibration target range distribution map.

[0029] Before performing K-means clustering on remote sensing images, the optimal number of cluster centers K is automatically determined by the elbow method. After unsupervised learning classification of the images, a distribution map with K categories is formed. After basic morphological operations are performed on the distribution map, each category is analyzed to extract the target range distribution map.

[0030] The beneficial effects of the present invention compared with the prior art are as follows:

[0031] (1) Comprehensively consider the delineation standards of remote sensing radiation calibration target ranges, take the image spatial spectral distribution characteristics as the detection basis, combine the long time series of images with signal domain decomposition analysis, and combine long time series analysis with empirical mode decomposition. It is possible to use multi-source long time series remote sensing images for intelligent extraction of remote sensing radiation calibration target ranges. The overall process is scientific and can be used for intelligent extraction over a large range of terrain.

[0032] (2) The Mann-Kendall algorithm is used to detect long-term remote sensing images, detect changing pixels and quasi-stable pixels, and evaluate the spatial uniformity of time-series quasi-stable pixels, which is more reasonable and scientific than using single-time images for evaluation.

[0033] (3) The BEMD algorithm is used to extract the texture features of the image, and unsupervised learning clustering is finally performed as one of the image features. The clustering result is object-oriented, and the extracted target area is more complete and can fully reflect the characteristics of the site. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] The present invention will be further described below in conjunction with the accompanying drawings:

[0035] Figure 1 The figure is a flow chart of the method of the present invention. DETAILED DESCRIPTION

[0036] like Figure 1As shown, the present invention provides a remote sensing radiation calibration target range extraction method integrating signal domain features, using remote sensing image comprehensive analysis to extract signal domain information and intelligently cluster remote sensing radiation calibration target ranges, comprehensively considering remote sensing radiation calibration target range delineation standards, taking image spatial spectrum distribution features as the detection basis, combining long time series of images with signal domain decomposition analysis, thereby intelligently extracting remote sensing radiation calibration target range distribution maps, including the following steps:

[0037] A. Extract spatial-spectral distribution features of multi-source time-series remote sensing images, including the following analysis contents:

[0038] a1) Feature band selection analysis: By assigning weights to the original bands of the image, the feature bands are selected. The specific implementation steps are as follows:

[0039] When evaluating the image band information, the factors to be considered are the amount of information in the band and the noise contained in the band. For the information evaluation of the image, the grayscale distribution is used to describe its size. The grayscale histogram is used to statistically evaluate the grayscale distribution of the image and reflect the amount of information. In the present invention, the input of the algorithm is the original image after the preprocessing steps (radiation calibration, atmospheric correction and other basic remote sensing image processing processes) at the same time and geographical distribution. The grayscale statistical process can be summarized as the following formula:

[0040]

[0041] Where: P i It refers to the probability of the i-th gray level appearing. The number of gray levels n can be customized to different values ​​according to the selected area. It is recommended to normalize the gray values ​​of the original image pixels after preprocessing and divide the gray values ​​into 20 levels for subsequent processing. H is the amount of information in the image band.

[0042] When selecting the calibration target range, the interference of image quality should also be considered. The local standard deviation noise evaluation method is used to evaluate the noise of each band. The main steps are:

[0043] Cut the single-band grayscale image into 8×8 regions and calculate the local standard deviation std and mean mean of each region.

[0044]

[0045] Where: num is the total number of regions after cutting, gray i is the grayscale value of the ith region.

[0046] Take a suitable equidistant segmentation interval between the local standard deviation maximum values. The number of segments is generally about 150. Count the situations where the previously calculated local standard deviation falls into each interval. Select the interval with the largest number of local standard deviations to obtain the average value, and use the average value as the noise value of the entire image.

[0047] Calculate the grayscale mean of the entire image, and use the ratio of the noise value to the grayscale mean as the noise estimation value N of the image i .

[0048] The weight that should be given to each band can be calculated by weighted superposition of the information content and noise estimation value of the image band. It should be noted that the blue band is seriously polluted by the atmosphere and is not suitable as the band selected for the target range, so its weight is 0. For other bands, the weight can be calculated according to the formula:

[0049]

[0050] In the formula: BI i is the weight of band i, n is the band index, H i is the information content of the i-band image, H min is the minimum information content of the image, H max is the maximum amount of information in the image, N min is the minimum noise estimate of the image, N max is the maximum noise estimate of the image, where the sum of all band weights is 1.

[0051] The selection of the shooting range needs to meet certain site characteristics, including spatial uniformity, temporal consistency, and being away from water bodies and buildings. The specific steps are as follows:

[0052] a2) Spatial uniformity analysis: Spatial uniformity is determined by using Getis statistic or CV index. For example, Getis index is used to determine: is the Getis statistic of pixel (i, j), T is the interval in which the Getis statistic meets the requirements, The evaluation of spatial uniformity can be completed by calculating the Getis statistic of the pixel.

[0053] a3) Temporal consistency (stability) analysis: Temporal consistency means that on a single pixel, each characteristic attribute on the pixel is guaranteed to be stable in time series. The specific mathematical calculation method is to perform statistics on the mean and standard deviation of the gray value of the time series image. After meeting the requirements (generally based on the condition of three times the mean error), the pixel meets the temporal consistency characteristics. The specific process can be described as the following formula:

[0054]

[0055] Where: is the reflectivity value of the pixel (i, j) in the image at the tth moment (0<t≤T), is the mean reflectivity of the pixel (i, j) in the image at all times, and σ(i, j) is the standard deviation of the reflectivity of the pixel (i, j) in the image at all times.

[0056] a4) Geographical distribution characteristics analysis: In order to ensure that the calibration range has little change with the season, the calibration range should have little or no vegetation coverage. Geographical distribution characteristics analysis further includes: Normalized vegetation index analysis and normalized building index analysis.

[0057] The Normalized Difference Vegetation Index (NDVI) is used to determine whether the area has low vegetation coverage. is the NDVI value of the pixel at position (x, y) at time t, NIR is the pixel value at the corresponding position of the near infrared band of the image, R is the pixel value at the corresponding position of the red band, and w NDVI The threshold for the maximum NDVI value that determines whether there is little or no vegetation in the area.

[0058]

[0059] If the above conditions are met, it is considered that the vegetation coverage in the area is low, so it is an important characteristic attribute for shooting range selection.

[0060] If it is close to a densely populated area or industrial facility area, human daily activities will affect or damage the selected shooting range. Therefore, the selected calibration shooting range should be far away from densely populated areas and densely built areas. The Normalized Difference Built-up Index (NDBI) is used to determine whether the area is densely populated or densely built. is the NDBI value of the pixel at position (x, y) at time t, SWIR is the pixel value at the corresponding position of the short-wave infrared band of the image, NIR is the pixel value at the corresponding position of the near-infrared band, and w NDBI The threshold for the maximum NDBI value is set to determine whether the area is densely populated and densely built.

[0061]

[0062]

[0063] If the above conditions are met, the population density or building density in the area is considered to be low, which is also an important statistical characteristic for shooting range selection.

[0064] B. Preprocessing and texture feature extraction of multi-source time series remote sensing images, including the following analysis:

[0065] b1) Before performing two-dimensional empirical mode decomposition on multi-source time series remote sensing images, preprocessing must be performed first. Mann-Kendall is a non-parametric statistical test method. Its advantages are that it does not require the measured values ​​to follow a normal distribution, nor does it require the trend to be linear, and it is not affected by missing values ​​and outliers. It has been widely used in trend significance tests for long time series data. The Mann-Kendall algorithm is used to detect the changing pixels and quasi-stable pixels at each pixel position in the long time series, and the changing pixels are removed while the average quasi-stable pixels are superimposed to construct a stable pixel image of a single time node, forming the basic data for BEMD decomposition.

[0066] Since spatial uniformity analysis needs to consider the stability of long-term time series pixels, the spatial uniformity analysis is performed on the multi-source time series remote sensing images preprocessed by the Mann-Kendall algorithm, while the temporal consistency (stability) analysis and geographic distribution characteristic analysis are performed on the original multi-source time series remote sensing images.

[0067] b2) Texture feature analysis based on two-dimensional empirical mode decomposition. By performing two-dimensional empirical mode decomposition (BEMD) on remote sensing images, we can obtain several intrinsic mode functions (IMFs), which correspond to three parts of information: high-frequency IMF corresponds to noise information, medium-frequency IMF corresponds to detail information, and low-frequency IMF corresponds to trend information. The comprehensive use of these texture feature information is conducive to the further selection of the target range.

[0068] Two-dimensional empirical mode decomposition adaptively decomposes remote sensing images into multiple intrinsic signal functions and a residual signal. The purpose is to mine the depth information of remote sensing images using signal science methods, and focus on the high-frequency and low-frequency parts of the decomposed signal components. The high-frequency part contains information such as noise, and the low-frequency part contains information such as distribution trends. These information are analyzed, the high-frequency components containing noise are removed, and other intrinsic signal functions are weighted and superimposed with the residual signal to form the basic supplementary information for unsupervised learning classification. The traditional remote sensing image target range selection method focuses on the spectral information of the image, while ignoring the role of the image texture information in the target range selection. BEMD is used to mine these hidden information and fuse it with the spectral information. The comprehensive use of the spectral information and texture information of the remote sensing image target range can improve the accuracy of intelligent target range extraction. Feature fusion refers to the superposition of the extracted spatial spectrum distribution features and texture features to form basic feature information.

[0069] For texture information, different remote sensing images applied in the BEMD algorithm will adaptively decompose different numbers of intrinsic mode functions, but their basic meanings are the same. The first decomposed IMF generally contains a lot of noise information, and the subsequent IMFs contain texture, seasonal trend and other information. In order to rationally utilize this information, in order to provide a more accurate range selection process algorithm, and to balance the importance of information in the subsequent feature band selection and construction, the design extracts the first three IMF components in order except for the first IMF. If it is insufficient, the last IMF is copied for supplementation. The overall weight of the three IMFs is 3, and the corresponding weight recommendation values ​​are 1.5, 1, and 0.5 in order, respectively. In this way, the neighborhood and trend characteristics of the image can be better utilized after removing the noise interference of the original image. The constructed feature band can be expressed as follows:

[0070] IM i =IMF i ;

[0071] IM=1.5IM 1 ∪IM 2 ∪0.5IM 3 ;

[0072] Where: IM is the constructed characteristic band group, there are three bands IM 1 ,IM 2 ,IM 3 , the above formula ∪ represents that it is composed of three single bands with different weights.

[0073] C. Extract the target range of unsupervised learning classification based on the extracted features of multi-source time series remote sensing images. The extraction of the target range based on unsupervised learning classification should integrate the previously extracted spatial spectrum distribution features and texture features, superimpose the original image with these features, and complete the comprehensive consideration of the features through the constructed weights to form a feature remote sensing image with the number of bands being the sum of the number of features and the number of bands of the original image. The feature images are clustered using the K-means clustering method, and the clustering results are analyzed after the morphological operation to obtain the final remote sensing radiation calibration target range distribution map. The specific analysis includes the following:

[0074] c1) Analysis of the optimal number of clusters: Before performing K-means classification on remote sensing images, the optimal number of cluster centers K can be automatically determined by the elbow method; after unsupervised learning classification of the images, a distribution map with K categories is formed. After performing basic morphological operations such as corrosion and expansion, each category is analyzed to extract the target range distribution map.

[0075] c2) Analysis of feature clustering results: The final target range distribution map can be obtained through analysis. The main characteristics considered in the analysis process are still the high temporal and spatial consistency and good geographical distribution characteristics. After analyzing the characteristics of the cluster center and the spectral curve, the optimal classification cluster can be obtained, which is the final extracted target range distribution.

[0076] The present invention integrates the basic geographical characteristics of the remote sensing radiation calibration target range and the texture characteristics of the target range in spatial distribution. By combining long time series analysis with empirical mode decomposition, it can use multi-source long time series remote sensing images to perform intelligent extraction of the remote sensing radiation calibration target range. The overall process is scientific and comprehensive factors are considered.

[0077] The present invention adopts the Mann-Kendall algorithm to detect the changing pixels and quasi-stable pixels at the position of each pixel in a long time series, and determines the stable image by eliminating the changing pixels and averaging the quasi-stable pixels, so as to determine the optimal analysis unit in the texture feature analysis link.

[0078] The present invention adopts multi-source time-series remote sensing image data for comprehensive analysis, refers to the basic characteristics of target range selection at home and abroad, comprehensively analyzes the relationship between target range selection criteria, and proposes a set of scientific and reasonable intelligent selection processes based on the spatiotemporal distribution characteristics and texture characteristics of target range selection.

[0079] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein by equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A remote sensing radiation calibration target range extraction method integrating signal domain features, characterized by: The following steps are involved: A: Extract spatial-spectral distribution features of multi-source time-series remote sensing images, including: a1) Characteristic band selection and analysis; a2) Spatial uniformity analysis; a3) Time consistency analysis; a4) Analysis of geographical distribution characteristics; B: The Mann-Kendall algorithm is used to preprocess the multi-source time series remote sensing images before texture feature extraction. Then, the texture features of the preprocessed multi-source time series remote sensing images are extracted based on two-dimensional empirical mode decomposition, and the texture features are fused with the spatial spectrum distribution features to obtain the basic feature information of the multi-source time series remote sensing images. C: Based on the fused basic feature information, unsupervised classification learning is performed on multi-source time series remote sensing images to obtain the target range distribution map.

2. The method for extracting a remote sensing radiation calibration target range by integrating signal domain features according to claim 1, characterized in that: In step a1), the characteristic bands are selected by assigning weights to the original bands of the image, wherein the weights are calculated by weighted superposition of the information content of the image bands and the noise estimation value of the image.

3. The method for extracting a remote sensing radiation calibration target range by integrating signal domain features according to claim 2, characterized in that: The information content of the image band is obtained through grayscale statistics; The noise estimation value of an image is the ratio of the noise value of the entire image to the grayscale mean.

4. The method for extracting a remote sensing radiation calibration target range by integrating signal domain features according to claim 3 is characterized in that: The band weight is calculated as follows: In the formula: BI i is the weight of band i, n is the band index, H i is the information content of the i-band image, H min is the minimum information content of the image, H max is the maximum amount of information in the image, N min is the minimum noise estimate of the image, N max is the maximum noise estimate of the image, and the sum of all band weights is 1; The blue band is seriously polluted by the atmosphere, so its weight is 0. For other bands, the weights are calculated according to the above formula.

5. The remote sensing radiation calibration target range extraction method integrating signal domain features according to claim 1 is characterized in that: In step a2), Getis statistic or CV index is used to judge the spatial uniformity of the multi-source time series remote sensing images preprocessed by the Mann-Kendall algorithm.

6. The method for extracting a remote sensing radiation calibration target range by integrating signal domain features according to claim 1, characterized in that: In step a3), the temporal consistency of the images is determined by statistically analyzing the mean and standard deviation of the grayscale values ​​of the time-series images.

7. The method for extracting a remote sensing radiation calibration target range by integrating signal domain features according to claim 1, characterized in that: The geographical distribution characteristics analysis in step a4) further includes: normalized vegetation index analysis and normalized building index analysis.

8. The method for extracting a remote sensing radiation calibration target range by integrating signal domain features according to claim 1, characterized in that: In step B, the texture features of the preprocessed multi-source time series remote sensing images are extracted based on two-dimensional empirical mode decomposition. The specific implementation steps are as follows: The intrinsic mode function (IMF) of remote sensing images is obtained by two-dimensional empirical mode decomposition. The three IMF components except the first IMF are extracted. If the components are insufficient, the last IMF component is copied to supplement them. The overall weight of the three IMF components is 3.

9. The method for extracting a remote sensing radiation calibration target range by integrating signal domain features according to claim 1, characterized in that: The specific implementation steps of step C are as follows: The extraction of the target range based on unsupervised learning classification requires the integration of the spatial uniformity, temporal consistency, geographical distribution characteristics and texture characteristics extracted previously, superimposing the original image and the extracted features, and completing the comprehensive consideration of the features through the constructed weights to obtain the characteristic remote sensing image. The characteristic remote sensing images are clustered using the K-means clustering method, and the clustering results are analyzed after the morphological operation to obtain the final remote sensing radiation calibration target range distribution map.

10. The method for extracting a remote sensing radiation calibration target range by integrating signal domain features according to claim 9, characterized in that: Before performing K-means clustering on remote sensing images, the optimal number of cluster centers K is automatically determined by the elbow method. After unsupervised learning classification of the images, a distribution map with K categories is formed. After basic morphological operations are performed on the distribution map, each category is analyzed to extract the target range distribution map.

Images