CN115452167A - Satellite remote sensor cross calibration method and device based on invariant pixel - Google Patents

Abstract

The invention provides a satellite remote sensor cross calibration method and device based on an invariant pixel. The method comprises the following steps: determining a sequence of image pairs; obtaining the apparent reflectivity of a single pixel in each image pair and inputting the apparent reflectivity into an invariant pixel detection model to obtain an invariant pixel in each image pair; acquiring spectrum matching factors of a remote sensor to be calibrated and a reference remote sensor for spectrum matching, and determining the spectrum correction apparent reflectivity of the reference remote sensor of an invariant pixel in each image pair; and performing orthogonal regression on the spectrum correction apparent reflectivity of the reference remote sensor of the invariant pixel in each image pair and the apparent reflectivity of the corresponding remote sensor to be calibrated to determine a cross calibration coefficient. The invention automatically detects an invariant pixel target in a scene, calculates a spectrum matching factor to correct the spectrum response difference between two sensors, establishes a linear fitting relation of the apparent reflectivity of an invariant pixel sample, and obtains a cross calibration result and a long-time sequence cross calibration coefficient thereof.

Description

Method and device for cross-calibration of satellite remote sensor based on invariant pixels

technical field

The invention relates to the technical field of remote sensor calibration, in particular to a satellite remote sensor cross-calibration method and device based on invariant pixels.

Background technique

With the rapid development of informatization and globalization and the continuous advancement of aerospace remote sensing detection technology, remote sensing applications have gradually penetrated into various fields of human activities. Accurate calibration of remote sensors is an important prerequisite for quantitative remote sensing applications.

There are many radiation calibration methods for remote sensors: in absolute radiation calibration, the absolute radiation value of the observed target needs to be obtained in advance, the implementation process is difficult and difficult to apply to historical satellite data recalibration; on-orbit satellite calibration method Relying on on-board calibration equipment, many satellites lack on-board calibration equipment or are limited by the development level of on-board calibration equipment, the calibration accuracy is low or on-orbit satellite calibration cannot be performed.

Cross-calibration is an effective and easy on-orbit alternative calibration method, which takes the sensor with high absolute calibration accuracy as the reference, and performs cross-calibration on the sensor to be calibrated by observing the same target at the same time. However, the existing cross-calibration methods have the disadvantages of difficulty in selecting an invariant target, single surface characteristics of the observation target, small dynamic range of reflectivity, and low calibration frequency.

Contents of the invention

The present invention provides a satellite remote sensor cross-calibration method and device based on invariant pixels, which are used to solve the problems of difficult selection of invariant targets, small dynamic range of target reflectivity, and problems in cross-calibration of remote sensors in the prior art. The defect of low calibration frequency. The invention does not need to manually select an invariant target, and can effectively increase the frequency of cross-calibration between sensors. At the same time, a large number of invariant pixels effectively increases the coverage of reflectivity, and is applicable to large scene data.

The invention provides a satellite remote sensor cross-calibration method based on invariant pixels, comprising:

Determining a sequence of image pairs, wherein each image pair includes multi-spectral images obtained by observing the same observation scene separately by the remote sensor to be calibrated and the reference remote sensor at the same time on the same day;

Obtain the apparent reflectance of a single pixel in each image pair and input the invariant pixel detection model to obtain the invariant pixel in each image pair;

Obtain the spectral matching factors of the remote sensor to be calibrated and the reference remote sensor, perform spectral matching on the reference remote sensor apparent reflectance of the constant pixel in each image pair based on the spectral matching factor, and determine each The spectrally corrected apparent reflectance of the reference sensor for the unchanged pixel in the image pair;

Orthogonal regression is performed on the spectrally corrected apparent reflectance of the reference remote sensor of the constant pixel in each image pair and the corresponding apparent reflectance of the remote sensor to be calibrated to determine the cross-calibration coefficient.

According to a method for cross-calibrating satellite remote sensors based on invariant pixels provided by the present invention, the determination of the sequence of image pairs includes:

Collect the multi-spectral images obtained from the observations of the same observation scene by the remote sensor to be calibrated and the reference remote sensor at the same time on the same day when the satellite passes through the border at the same time;

Preprocessing is performed on the multispectral image collected by the remote sensor to be calibrated and the multispectral image collected by the reference remote sensor to obtain a target image pair.

The image pair sequence is determined based on the target image pair collected over multiple days.

According to a satellite remote sensor cross-calibration method based on invariant pixels provided by the present invention, the preprocessing includes resolution resampling, rasterization and elimination of invalid pixels.

According to a method for cross-calibrating satellite remote sensors based on invariant pixels provided by the present invention, the elimination of invalid pixels includes:

Eliminate cloud pollution pixels, water body target pixels, and pixels with satellite observation zenith angles greater than or equal to 30°.

According to a satellite remote sensor cross-calibration method based on invariant pixels provided by the present invention, the invariant pixel detection model is established based on an iterative weighted multivariate change detection IR-MAD method.

According to a satellite remote sensor cross-calibration method based on invariant pixels provided by the present invention, the apparent reflectance of a single pixel in each image pair is acquired and input into the invariant pixel detection model, and the obtained Describe the invariant pixels of each image pair, including:

Obtaining the detection result of a single pixel in each image pair and calculating the apparent reflectance of the pixel;

Input the pixel apparent reflectance of a single pixel in each image pair into an invariant pixel detection model, construct a MAD variable and construct the observation value of a single pixel in each image pair based on the MAD variable;

An invariant pixel for each image pair is determined based on the observed value of a single pixel in each image pair and an invariant probability decision threshold.

According to a method for cross-calibrating satellite remote sensors based on invariant pixels provided by the present invention, the acquisition of the spectral matching factors of the remote sensor to be calibrated and the reference remote sensor includes:

Obtaining the entrance pupil radiance of the remote sensor to be calibrated and the entrance pupil radiance of the reference remote sensor;

Based on the entrance pupil radiance of the remote sensor to be calibrated and the pupil radiance of the reference remote sensor, spectral matching factors of the remote sensor to be calibrated and the reference remote sensor are determined.

According to a satellite remote sensor cross-calibration method based on invariant pixels provided by the present invention, the spectrally corrected apparent reflectance of the reference remote sensor of the invariant pixels of each image pair is compared with the corresponding image pair Orthogonal regression is performed on the apparent reflectance of the remote sensor to be calibrated in the unchanged pixel to determine the cross-calibration coefficient, including:

Establishing a linear fitting relationship between the spectrally corrected apparent reflectance of the reference remote sensor of the constant pixel of each image pair and the apparent reflectance of the uncalibrated remote sensor of the corresponding image pair return;

Based on the slope and intercept of the orthogonal regression, the cross-calibration coefficients obtained were determined.

According to a kind of satellite remote sensor cross-calibration method based on invariant pixel provided by the present invention, the determination of the cross-scale coefficient also includes afterwards:

Carrying out long-time series cross-calibration for each channel of the long-time series data of the remote sensor to be calibrated;

Based on the results of the cross-calibration of the long-time series, cross-calibration coefficients for the long-time series are determined.

The present invention also provides a satellite remote sensor cross-calibration device based on invariant pixels, including:

The acquisition module is used to determine a sequence of image pairs, each of which includes multispectral images obtained by observing the same observation scene separately by the remote sensor to be calibrated and the reference remote sensor at the same time on the same day;

The invariant pixel detection module is used to obtain the apparent reflectance of a single pixel in each image pair and input the invariant pixel detection model to obtain the invariant pixel in each image pair;

The spectral matching module is used to obtain the spectral matching factor of the remote sensor to be calibrated and the reference remote sensor, and perform spectral matching on the reference remote sensor apparent reflectance of the constant pixel in each image pair based on the spectral matching factor , determining the spectrally corrected apparent reflectance of the reference remote sensor of the constant pixel in each image pair;

The regression module is used to perform orthogonal regression on the spectrally corrected apparent reflectance of the reference remote sensor of the constant pixel in each image pair and the corresponding apparent reflectance of the remote sensor to be calibrated to determine cross-calibration coefficient.

The satellite remote sensor cross-calibration method and device based on invariant pixels provided by the present invention is different from the traditional cross-calibration method, and does not rely on the invariant calibration field manually selected by satellite transit, and is not limited to ground synchronization The harsh conditions of observation can effectively increase the frequency of cross-calibration between sensors. The detection of the constant pixel target is automatic, and it has a positive effect on the processing of long-term series data and recalibration of historical satellite data. The obtained invariant pixel targets are spatially discontinuous. Although the spatial characteristics of the surface cannot be completely analyzed, a large number of invariant pixel samples effectively increase the coverage of reflectivity, and can be applied to large scene data. The problem of small reflectance range and single sample of the target site observed by the traditional cross-calibration method is improved. The calibration results can be used for the research application of subsequent quantitative inversion products of remote sensing data and the long-term monitoring of the radiation response characteristics of remote sensors.

Description of drawings

In order to more clearly illustrate the present invention or the technical solutions in the prior art, the accompanying drawings that need to be used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the accompanying drawings in the following description are the present invention. For some embodiments of the invention, those skilled in the art can also obtain other drawings based on these drawings without creative effort.

Fig. 1 is one of the schematic flow charts of the satellite remote sensor cross-calibration method based on invariant pixel provided by the embodiment of the present invention;

Fig. 2 is the second schematic flow diagram of the satellite remote sensor cross-calibration method based on the invariant pixel provided by the embodiment of the present invention;

Fig. 3 is a schematic diagram of FY-3B/VIRR, MERSI channel spectral response function and SCIAMACHY hyperspectral sample provided by the embodiment of the present invention;

Fig. 4 is a schematic diagram of the orthogonal regression result of the constant pixel apparent reflectance provided by the embodiment of the present invention;

Fig. 5 is the long-term schematic diagram of VIRR relative calibration slope provided by the embodiment of the present invention;

Fig. 6 is a schematic structural diagram of a satellite remote sensor cross-calibration device based on invariant pixels provided by an embodiment of the present invention.

detailed description

In order to make the purpose, technical solutions and advantages of the present invention clearer, the technical solutions in the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the present invention. Obviously, the described embodiments are part of the embodiments of the present invention , but not all examples. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without creative efforts fall within the protection scope of the present invention.

The existing cross-calibration methods are mainly divided into two categories, one is the SNO cross-calibration method, and its main steps are:

1) Obtain the geographic location and observation time of the intersection of the orbits of the two satellites through orbit forecasting;

2) Based on the time and geographical location of the track forecast intersection, the pixel matching of the observation data has been selected for cross calibration, including time, observation geometry, and spatial location matching;

3) Spectral matching is performed according to the spectral response difference of similar channels of two remote sensors;

4) Regression analysis to calculate the calibration coefficient.

The other is the site cross-calibration method, the main steps of which are:

1) Select a suitable unchanging site as the observation target;

2) Obtain the invariant site observation data matching the observation conditions of the two remote sensors and project them to the same geographic grid;

3) Obtain the simulated apparent reflectance or apparent radiance of the reference sensor using a radiation transfer model and spectral matching;

4) Regression analysis to calculate the calibration coefficient.

Based on the above two existing cross-calibration methods, the following three cross-calibration problems can be concluded:

It is difficult to choose a constant target. SNO cross-calibration needs to use the orbit prediction model to select satellite orbit intersections, and the available cross-calibration data can only be obtained after strict pixel matching in limited intersection points. Site cross-calibration also requires manual selection and delineation of unchanged sites. , consume manpower and material resources.

The surface characteristics of the observation target are single, and the dynamic range of reflectivity is small. For example, using the desert invariant field for site cross-calibration, the visible light band reflectance of the desert site is about 0.2-0.3, but the radiation performance of the remote sensor has a certain target dependence, that is, the calibration coefficient will vary with the target radiation. The dynamics change, so a large number of target samples are required to cover the dynamic range of the remote sensor as much as possible to obtain a high-precision calibration result.

Calibration frequency is low. Whether it is SNO cross-calibration or site cross-calibration, the frequency of cross-calibration is low due to the difficulty of manual selection of constant targets, and continuous radiation calibration cannot be carried out to achieve the goal of long-term detection of remote sensor radiation performance.

Based on the above problems, the embodiment of the present invention provides a method for cross-calibration of satellite remote sensors based on invariant pixels, which realizes cross-calibration of remote sensors and solves the difficulty of selecting invariant targets when performing cross-calibration of remote sensors. , The target reflectivity has a small dynamic range and a low calibration frequency.

Below in conjunction with Fig. 1-Fig. 5 describe the satellite remote sensor cross-calibration method based on invariant pixel of the present invention, as shown in Fig. 1, method at least comprises the following steps:

Step 101. Determine the sequence of image pairs, wherein each image pair includes multispectral images obtained by observing the same observation scene at the same time on the same day and at the same time by the remote sensor to be calibrated and the reference remote sensor;

Step 102. Obtain the apparent reflectance of a single pixel in each image pair and input it into the invariant pixel detection model to obtain the invariant pixel in each image pair;

Step 103, obtain the spectral matching factor of the remote sensor to be calibrated and the reference remote sensor, perform spectral matching on the reference remote sensor apparent reflectance of the constant pixel in each image pair based on the spectral matching factor, and determine the The spectrally corrected apparent reflectance of the reference remote sensor for the constant pixel in each image pair;

Step 104 , performing orthogonal regression on the spectrally corrected apparent reflectance of the reference remote sensor of the constant pixel in each image pair and the corresponding apparent reflectance of the remote sensor to be calibrated to determine a cross-calibration coefficient.

Regarding step 101, it should be noted that the sequence of image pairs includes multiple image pairs, and each image pair is a simultaneous phase multi-spectral image of the same target scene collected by two remote sensors when a satellite passes through the border at the same time. The two image pairs may be two sets of multispectral images collected at the same time for the same observation scene in two adjacent days, or two groups of multispectral images collected for the same observation scene at different times on the same day, or It can be two sets of multispectral images with different date, collection time, and observation scene. The image pair sequence needs to collect satellite data for several consecutive days, and the continuous collection time must be within the satellite orbit return period.

In addition, each multispectral image contains satellite data of multiple channel bands. When selecting an image pair, it is necessary to perform channel matching on two multispectral images in the same phase and for the same target scene. One to one correspondence.

With regard to step 102, it should be noted that the detection result of a single pixel in each image pair is generally the radiometric value of the remote sensing sensor, that is, the DN count value. The apparent reflectance of each pixel can be calculated based on the detection results. The invariant pixel detection model is a pre-established model that can filter out invariant pixels from the input image pair according to the apparent reflectance. For each single-day image pair, after passing through the invariant pixel detection model, the invariant pixel of the image pair can be output, and after inputting the sequence of image pairs into the invariant pixel detection model in turn, you can get the corresponding A sequence of unchanged cells.

For step 103, it should be noted that due to the difference in the spectral response of the corresponding channel between the remote sensor to be calibrated and the reference remote sensor, that is, for the same entrance pupil radiation, the two remote sensors will obtain different measured values, so it is necessary to Spectral matching is performed on corresponding channels to correct calibration errors due to differences in spectral response.

In this embodiment, the spectral matching factor of the remote sensor to be calibrated and the reference remote sensor needs to use the observation data of the hyperspectral instrument as a hyperspectral sample, which is respectively convoluted with the channel spectral response function of the remote sensor to be calibrated and the reference remote sensor, to obtain The entrance pupil radiance of the corresponding remote sensor can be obtained by establishing the relationship between the entrance pupil radiance of the matching channel of the remote sensor to be calibrated and the reference remote sensor.

Regarding step 104, it should be noted that any image pair with invariant pixels obtained by the invariant pixel detection model can obtain cross calibration coefficients after processing in steps 103 and 104. However, due to the wide range of scenes used for statistical analysis, the results of single-day image pairs only contain part of the area information in the scene that meets the observation zenith angle constraints. In order to include as many unchanging pixel samples as possible in a single regression, In this embodiment, the sequence of image pairs is collected, and the invariant pixels detected by consecutive days are combined for fitting processing. Based on the short-term stability of the radiation response of the remote sensor, this combination process is feasible, and will help to expand the range of reflectance values of the pixel samples, and effectively improve the regression effect and calibration accuracy.

The satellite remote sensor cross-calibration method based on invariant pixels in the embodiment of the present invention is aimed at the problem of invariant target selection in the existing cross-calibration method, the problem of limited dynamic range of the reflectivity of the observed target, and the long-term sequence of remote sensing. An effective solution is proposed for the problem of continuous calibration monitoring of the instrument. This method is different from the traditional cross-calibration method. It does not rely on the constant calibration field manually selected by satellite transit, and is not limited by the harsh conditions of ground synchronous observation. It can effectively increase the frequency of cross-calibration between sensors. The detection of the constant pixel target is automatic, and it has a positive effect on the processing of long-term series data and recalibration of historical satellite data. The obtained invariant pixel targets are spatially discontinuous. Although the spatial characteristics of the surface cannot be completely analyzed, a large number of invariant pixel samples effectively increase the coverage of reflectivity, and can be applied to large scene data. The problem of small reflectance range and single sample of the target site observed by the traditional cross-calibration method is improved. The calibration results can be used for the research application of subsequent quantitative inversion products of remote sensing data and the long-term monitoring of the radiation response characteristics of remote sensors.

It will be appreciated that determining the sequence of image pairs includes:

Collect the multi-spectral images obtained from the observations of the same observation scene by the remote sensor to be calibrated and the reference remote sensor at the same time on the same day when the satellite passes through the border at the same time;

The multispectral image collected by the remote sensor to be calibrated and the multispectral image collected by the reference remote sensor are preprocessed to obtain the target image pair.

Based on the target image pair acquired over multiple days, an image pair sequence is determined.

It should be noted that due to the different spatial resolutions between different remote sensors and different channels, and the different data dimensions, subsequent statistical analysis cannot be carried out. Therefore, the multispectral images collected by the remote sensors to be calibrated and the reference The multispectral images collected by the remote sensor are preprocessed, and the preprocessed image pair is the target image pair. Collect target image pairs in consecutive days to form an image pair sequence.

Understandably, preprocessing includes resolution resampling, rasterization, and removal of invalid cells.

It should be noted that during preprocessing, the same channel bands of the two multispectral images in the same image pair need to be matched one by one, and the basis of subsequent steps is the image pair under the same channel band, and each The DN values of the two multispectral images of the image pair under each channel band. In the embodiment of the present invention, the channel band matching can be realized by performing resolution resampling on the satellite data of each image pair. Project the satellite remote sensing data corresponding to the multispectral image to the same geographic grid, and organize the rasterized satellite remote sensing data into a satellite dataset. in. Geographic grid is a data form that divides geographic space into regular grids, each grid is called a unit, and assigns corresponding attribute values to each unit to represent satellite remote sensing data.

In the DN values of different channels obtained by the two remote sensors, there are invalid data. Since the invariant pixel detection model is developed based on statistical principles, the dimensions of the two sets of input data are required to be consistent, so the invalid pixels in the two images in each image pair need to be correspondingly eliminated.

Understandably, the removal of invalid cells includes:

Eliminate cloud pollution pixels, water body target pixels, and pixels with satellite observation zenith angles greater than or equal to 30°.

It should be noted that when the invariant pixel detection model is used to detect invariant pixels in the scene, the changed pixels will be gradually eliminated in the iterative process of the algorithm. Pixels, also known as cloud pollution pixels, such as clouds, dust and other objects should be eliminated before the algorithm corresponding to the invariant pixel detection model runs. Although areas such as water bodies and oceans have relatively stable reflectance characteristics, their reflectance spectrum characteristics are quite different, and the signals in some channel bands are weak. For water body target pixels, the land and ocean mask data in the data are used to eliminate them.

In addition, after removing invalid points, clouds, dust, and ocean water pixels from the image pair sequence data, there are still millions of pixel samples, among which there are a large number of pixel samples with too large zenith angles observed by satellites. The satellite observation zenith angle is too large, which will lead to the decrease of pixel spatial resolution and detection radiation accuracy. In order to improve the recognition accuracy of the invariant pixel detection model for the invariant pixel, the satellite observation zenith angle is less than 30°. for subsequent statistical analysis.

It can be understood that the invariant pixel detection model is established based on the iterative weighted multivariate change detection IR-MAD method.

It should be noted that with the advancement of computer information technology and the widespread application of statistical analysis methods in scientific research, remote sensing researchers have gradually applied some mathematical methods to the processing and analysis of satellite data. In the prior art, the difference image comparison method and the ratio method are suitable for the analysis of single-channel images, but not suitable for multi-channel satellite remote sensor data. The principal component analysis (PCA) method can synthesize the change information of each channel, but it cannot eliminate the correlation between different remote sensor channels. And a multivariate change detection technique (MAD) with linear scale invariance can eliminate the correlation within the same-phase sensor channel and the correlation between different channels of the sensor. Based on this, the embodiment of the present invention uses the IR-MAD method to construct an invariant pixel detection model applied to different remote sensors in the same phase observation scene, and applies it to the cross-calibration of the remote sensors.

It can be understood that, in the embodiment of the present invention, the invariant pixel detection model is constructed by the IR-MAD method, and the invariant pixel in each image pair is obtained. Among them, two n-channel multispectral images of the same scene collected by the remote sensor to be calibrated and the reference remote sensor at the same time are analyzed and processed using the invariant pixel detection model, at least including the following steps:

Step 201, obtaining the detection result of a single pixel in each image pair and calculating the apparent reflectance of the pixel;

Step 202, input the pixel apparent reflectance of a single pixel in each image pair into the invariant pixel detection model, construct a MAD variable and construct the observation value of a single pixel in each image pair based on the MAD variable;

Step 203, based on the observation value of a single pixel in each image pair and the invariant probability decision threshold, determine the invariant pixel of each image pair.

Regarding step 201, it should be noted that the apparent reflectance (TOA) refers to the reflectance of the top of the atmosphere, and its value is the sum of the contribution of the surface reflectance and the atmospheric transmittance. Using the fixed calibration coefficient at the initial stage of satellite operation, the apparent reflectance of the pixel is calculated from the DN value of the remote sensor channel:

Among them, slope _i and bias _i are the fixed calibration slope and intercept corresponding to the i-th channel of any remote sensor, DN _i is the DN value obtained by any remote sensor, d ² is the sun-earth distance correction factor, θ _s is the sun zenith angle.

Regarding step 202, it should be noted that after inputting the pixel apparent reflectance of a single pixel in each image pair into the invariant pixel detection model, it is first necessary to construct the MAD variable. The process of constructing MAD variables includes:

Step 2021, using vectors F=(F ₁ ...F _n ) ^T , G=(G ₁ ...G _n ) ^T to represent the apparent reflectance after conversion of DN values in two multispectral images in an image pair, for all The bands are linearly combined to obtain typical variables U and V, as shown in Formula 1:

Among them, n is the total number of matching channels of the current image pair, m and n are constant vectors, which can be obtained by solving the coupled generalized eigenvalue equation of formula 2:

Among them, ρ is the correlation coefficient of typical variables U and V, and ∑ _ff , ∑ _gg , ∑ _fg , ∑ _gf are the covariance matrices of image vectors F and G.

Step 2022, obtain the MAD variable based on the difference of typical variables U and V, as shown in formula 3:

Among them, MAD _i represents the MAD variable of the i-th channel of the remote sensor.

Step 2023, based on the linear scale invariance of the MAD variable, the observed value Z is obtained, as shown in formula 4:

为遥感器第i通道的MAD变量的标准差，Z为标准化MAD变量的平方和，近似为卡方分布(χ²)，有n个自由度，Z值越小，像元不变概率越高。in,

is the standard deviation of the MAD variable of the i-th channel of the remote sensor, and Z is the sum of squares of the standardized MAD variable, which is approximately a chi-square distribution (χ ² ), with n degrees of freedom, and the smaller the Z value, the higher the probability of pixel invariance .

Regarding step 203, it should be noted that, for two n-channel multispectral images of the same scene collected at the same time by two remote sensors in any image pair, ideally the difference between them is caused by the instrument response itself and noise, atmospheric Caused by random effects such as fluctuations, from the central limit theorem, MAD variables conform to the ideal normal distribution. However, the MAD variables associated with change observations will more or less deviate from this multivariate normal distribution. In the presence of change, the focus is on successive iterations to establish a better and better no-change background, then MAD Transform sensitivity will be improved.

Since a single MAD transformation is difficult to achieve the ideal invariant pixel detection effect, when estimating the sample mean and covariance matrix, the embodiment of the present invention weights the sample data through the invariant probability determined by a single iteration, giving more The pixels with high invariant probability have greater weights, and the MAD variables of the next iteration are determined through canonical correlation analysis, and multiple iterations are used to obtain better detection results of invariant pixels.

Specifically, step 203 includes the following sub-steps:

Step 2031, based on the observed value, calculate the constant probability weight value;

It should be noted that, for each iteration, the constant probability weight value Pr can be determined by the chi-square distribution test result formula 5 of the observed value Z:

For the IR-MAD algorithm, it is necessary to set three iteration stop thresholds, namely, the change threshold of the typical variable correlation coefficient, the maximum number of iterations, and the minimum number of NCPs, so that it stops iterative processing of an image pair when the algorithm is optimally converged and when the low-probability detection fails. , to ensure the automatic operation of the algorithm. When the change difference of the two canonical variable correlation coefficients is less than the canonical variable correlation coefficient change threshold, the algorithm can be regarded as converged and the iteration is stopped; at the same time, in order to ensure that there are enough unchanged pixel samples for subsequent cross-calibration regression analysis , it is necessary to set the minimum threshold of the number of NCPs. When the number of NCPs is less than this value after a certain iteration, the iteration stops.

Step 2032 , setting the invariant probability decision threshold, and determining the invariant pixel of each image pair based on the observed value of a single pixel in each image pair and the invariant probability decision threshold.

可用于后续交叉定标分析。It should be noted that, since this embodiment seeks to find an invariant pixel with a greater probability, when the number of matching channel groups n is determined, test and analyze multiple invariant probability confidences, and comprehensively consider the operating efficiency of the algorithm and the invariant To change the detection effect of the pixel, select the appropriate invariant probability decision threshold t, when the observed value Z<t, the confidence that the pixel sample is considered to be the invariant pixel is higher than

Can be used for subsequent cross-calibration analysis.

Therefore, in the present embodiment, the samples satisfying the condition of formula 6 are selected as invariant pixels (NCPs) in the pixel samples:

P为卡方检验值小于等于t的概率。Among them, t is the upper quantile of the chi-square distribution with n degrees of freedom, that is, the probability decision threshold of the constant pixel,

P is the probability that the chi-square test value is less than or equal to t.

In addition, the spatial location of NCPs selected by the statistical properties of the data without prior knowledge of the surface corresponds to features that are invariant between image pairs, but the location of NCPs changes with the difference in radiometric information between different image pairs.

It can be understood that obtaining the spectral matching factors of the remote sensor to be calibrated and the reference remote sensor includes:

Obtain the entrance pupil radiance of the remote sensor to be calibrated and the entrance pupil radiance of the reference remote sensor;

Based on the entrance pupil radiance of the remote sensor to be calibrated and the pupil radiance of the reference remote sensor, the spectral matching factor of the remote sensor to be calibrated and the reference remote sensor is determined.

It should be noted that in the embodiment of the present invention, the observation data of the hyperspectral instrument is used as the hyperspectral sample, which is respectively convolved with the channel spectral response function of the remote sensor to be calibrated and the reference remote sensor to obtain the entrance pupil radiance of the corresponding remote sensor, as shown in the formula 7 shows:

Among them, R _h is the radiance of the hyperspectral sample, _{fi_sensor} is the spectral response function of channel i of the hyperspectral instrument, and R _{i_sensor} is the entrance pupil radiance of the remote sensor to be calibrated or the reference remote sensor after convolution.

Under the same geometrical conditions of surface, atmosphere and observation time, the ratio of the incident radiance of the remote sensor to be calibrated and the matching channel of the reference remote sensor is called the spectral matching factor (SBAF). Establish the relationship between the entrance pupil brightness of the matching channel of the remote sensor to be calibrated and the reference remote sensor, as shown in Equation 8, and obtain the SBAF coefficient of the corresponding channel of the two remote sensors:

R _{i_CAL} = A _{i, j} × R _{j_REF} + B _{i, j} Formula 8

Among them, R represents the radiance, A and B are the spectral matching factors, i and j are the matching channel numbers of the remote sensor to be calibrated (CAL) and the reference remote sensor (REF), respectively, A _{i, j} and B _{i, j} represent the calibration The spectral matching factor of the remote sensor i channel and the reference remote sensor j channel is calculated by least squares regression fitting to calculate the SBAF of each matching channel.

It can be understood that the spectrally corrected apparent reflectance of the reference remote sensor of the constant pixel of each image pair is orthogonally regressed with the apparent reflectance of the uncalibrated remote sensor of the corresponding image pair of constant pixel, Determine cross-scaling factors, including:

Establish a linear fitting relationship between the spectrally corrected apparent reflectance of the reference remote sensor of the constant pixel of each image pair and the apparent reflectance of the remote sensor to be calibrated of the corresponding image pair of constant pixel to perform orthogonal regression;

Based on the slope and intercept of the orthogonal regression, the cross-calibration coefficients obtained were determined.

It should be noted that, for the invariant pixels acquired by IR-MAD detection, the apparent reflectance detection result ρ _{j_REF} of channel j of the reference remote sensor is converted using the spectral matching factor to obtain the spectrally corrected apparent reflectance under the spectral response of channel i The reflectance spectrum corrects the apparent reflectance ρ _{i_REF} , as shown in Equation 9:

ρ _{i_REF} =A _i,j ×ρ _{j_REF} +B _i,j Equation 9

For any image pair of invariant pixels obtained by IR-MAD detection, the spectrally corrected apparent reflectance ρ _{i_REF} of the i channel obtained through spectral matching and the apparent reflectance detection of the remote sensor to be calibrated As a result, ρ _{i_CAL} establishes a linear fitting relationship for orthogonal regression, as shown in Equation 10, so as to obtain the cross-calibration coefficient:

ρ _{i_REF} ＝a×ρ _{i_CAL} +b Formula 10

Among them, a is the calibration slope, and b is the calibration intercept. ρ _{i_CAL} is the apparent reflectance calculated based on the fixed calibration slope at the initial launch stage, so the calibration coefficients a and b are relative calibration results compared to the fixed calibration coefficients at the initial launch stage.

It can be understood that after determining the cross-calibration coefficient, it also includes:

Carrying out long-time series cross-calibration for each channel of the long-time series data of the remote sensor to be calibrated;

Based on the results of the cross-calibration of the long-time series, cross-calibration coefficients for the long-time series are determined.

It should be noted that the long-term cross-calibration of each channel of the remote sensor to be calibrated is used to monitor the long-term radiation response of the remote sensor channel and determine the trend of the long-term cross-calibration coefficient.

Provide the specific embodiment of the inventive method below:

Fengyun-3B (FY-3B) is the second of China's second-generation polar-orbiting meteorological satellites, carrying 11 remote sensing instruments, Visible Infrared Scanning Radiometer (VIRR) and Moderate Resolution Imaging Spectrometer (MERSI) are the two main instruments. FY-3B/VIRR has a total of 10 channel bands, with a spectral range of 0.43-12.5μm, including 7 visible near-infrared bands and 3 infrared emission bands, and the sub-satellite point spatial resolution is 1.1km. VIRR is mainly used to detect and identify cloud information, detect surface vegetation coverage, obtain relevant information on surface sea surface, and monitor the total amount of atmospheric water vapor. FY3B/MERSI has 19 reflective solar bands (0.41～2.13μm) and 1 thermal infrared band (11.25μm), and the sub-satellite point spatial resolution is 250m and 1000m. Channels 1, 2, 6, 7, 8, and 9 of VIRR are similar to channels 3, 4, 6, 1, 10_11, and 2 of MERSI (channel 8 of VIRR corresponds to channels 10 and 11 of MERSI). For specific indicators, see Table 1, so MERSI is used as the reference remote sensor, VIRR is used as the remote sensor to be calibrated, and a total of 6 groups of corresponding channels of VIRR are cross-calibrated.

As shown in Figure 2, this embodiment screens and preprocesses the image data pairs of the remote sensor VIRR to be calibrated and the reference remote sensor MERSI in the same phase, and linearly combines the matching channel data of the two remote sensors to construct the MAD variable. The IR-MAD invariant pixel detection model analyzes the invariant pixel samples in the scene. The obtained invariant pixels take the detection results of the reference remote sensor as the radiation reference, and after spectral matching and correction, they are consistent with the detection results of the remote sensor to be calibrated. Orthogonal regression was performed to obtain cross-scaling coefficients.

Table 1 FY3B/VIRR, MERSI spectral band indicators

It can be understood that, for step 101, the observation scene of the sequence of image pairs selected in this embodiment is Northwest China, and the center coordinates are 91° east longitude and 39° north latitude. The terrain is mainly plateaus, basins, and mountains, with an altitude of more than 1km. The landforms and landscapes are mainly the Loess Plateau, Gobi Desert, desert steppe and Gobi Desert. This area has a continental climate, influenced by the southern Himalayas, cold and dry in winter, hot in summer, with little precipitation, and drought is the main natural feature of this area. The seasonal variation of surface reflectance in the study area is small, which meets the ideal cross-calibration observation area standard of low aerosol content, low water vapor content and high probability of clear sky.

As shown in Table 1, there are differences in channel resolution between VIRR and MERSI. There are three spatial resolutions of 250m, 1km and 1.1km. The detection data of each channel is resampled to 1km spatial resolution and projected to the same geographic grid. The data size of a single scene is 1400*3200 pixels, and the time range is from January 21, 2011 to November 14, 2018.

It can be understood that, for step 101, in the DN values of the simultaneous phase multispectral images collected by VIRR and MERSI, there are some invalid points in each channel, and it is necessary to ensure that the data of each channel of each pixel used for statistical analysis are valid values , so cells with invalid values are discarded.

Since the significant change scene in the study area of the observation scene is mainly cloud, the L1 data of VIRR and MERSI have no cloud mask data, so the channel data of VIRR is used, and the threshold discriminant method is used for preliminary cloud identification. Firstly, the radiance ratios of channel 1 (0.58-0.68μm) and channel 2 (0.84-0.89μm) of VIRR can be used to effectively distinguish between clouds and clear sky areas. The radiation ratio of 1.64μm) can better distinguish the clouds above the ground from the objects with high reflectivity on the ground. For ocean water pixel targets, the land seamask data in VIRR and MERSI data are used to eliminate them. The scanning ranges of the field of view of VIRR and MERSI remote sensors are both ±55.4°, and the pixels with satellite observation zenith angles less than 30° are selected for subsequent statistical analysis.

It can be understood that for step 102, the fixed calibration coefficient of the remote sensor of the FY-3B satellite in the early stage of operation is used to calculate the apparent reflectance of the pixel from the channel DN value, and obtain the apparent reflectance image of the research scene in the same phase.

It can be understood that for step 102, for the two multispectral images of Northwest China collected by VIRR and MERSI at the same time, after data screening and preprocessing, 6 sets of matching channels of sample pixels for IR-MAD analysis are reserved Apparent reflectance data, ie n=6.

In this embodiment, the IR-MAD analysis is performed on data pairs of multiple groups of images, and is used for testing to select an appropriate iteration stop threshold parameter. After 20 times, the downward trend gradually becomes flat, and the algorithm tends to converge at 20-25 times. At this time, the change of the correlation coefficient of the typical variable between the two iterations is less than 0.001, so the difference between the two typical variable correlation coefficients can be set to be less than 0.001 When the algorithm converges, the iteration is stopped, and the maximum number of iterations of the algorithm is set to 30. At the same time, in order to ensure that there are enough unchanged pixel samples for subsequent cross-calibration regression analysis, the minimum number of NCPs is set to 400. When the number of NCPs is less than this value after a certain iteration, the iteration stops.

作为不变概率决策阈值，当观测值Z＜t＝1.635时，认为该像元样本为不变像元的置信度高于95％，可用于后续交叉定标分析。In this embodiment, it is necessary to obtain invariant pixels with an invariant probability of more than 90%. Probabilistic confidence is tested and analyzed, considering the efficiency of the algorithm and the detection effect of the invariant pixel, select

As the invariant probability decision threshold, when the observation value Z<t=1.635, the confidence that the pixel sample is considered to be an invariant pixel is higher than 95%, which can be used for subsequent cross-calibration analysis.

It can be understood that, for step 203, a relational expression between VIRR and MERSI matching channel entrance pupil brightness is established, and the SBAF coefficients of the corresponding channels of the two remote sensors are obtained, specifically including:

R _{i_VIRR} = A _{i, j} × R _{j_MERSI} + B _{i, j} Equation 11

R _{8_VIRR} = A _8,10 ×R _{10_MERSI} +A _8,11 ×R _{11_MERSI} +B _i,j Formula 12

Among them, R represents the radiance, A and B are the spectral matching factors, and i and j are the matching channel numbers of VIRR and MERSI, respectively. Formula 11 is used for the one-to-one matching channel of VIRR and MERSI, and formula 12 is used for channel 8 of VIRR and channels 10 and 11 of MERSI. The SBAF of each matching channel was calculated by least squares regression fitting, and the spectral matching factors of VIRR and MERSI were obtained as shown in Table 2.

Table 2 FY-3B VIRR and MERSI Spectral Matching Factor (SBAF)

It can be understood that, for step 203, for the invariant pixels obtained by IR-MAD detection, the apparent reflectance detection result ρ _{j_MERSI} of the MERSI j channel is converted by using the spectral matching factor to obtain the spectral correction under the spectral response of the i channel Apparent reflectance ρ _{i_MERSI} .

It can be understood that, for step 204, for any image pair obtained by IR-MAD detection, the i-channel spectral corrected apparent reflectance ρ _{i_MERSI} obtained through spectral matching and the apparent reflectance detection of VIRR As a result, ρ _{i_VIRR} establishes a linear fitting relationship as shown in Equation 12, so as to obtain the cross-calibration coefficient.

ρ _{i_MERSI} ＝a×ρ _{i_VIRR} +b Equation 12

Among them, a is the calibration slope, and b is the calibration intercept.

It should be noted that the nominal orbit return period of Fengyun-3 satellite is 5.5 days, and the orthogonal regression analysis is carried out on the combination of unchanged pixels detected and acquired for 5 consecutive days. Figure 4 shows the orthogonal regression results of the matching channels of each group of unchanged pixels from April 8 to 12, 2011. The linear fitting effect of the constant pixel in each channel is good, and the dynamic range of TOA is large.

Figure 5 shows the long-term trend of the relative calibration slope of each channel of VIRR from January 21, 2011 to November 14, 2018, which is used to monitor the long-term radiation response of the remote sensor channel. The method of the present invention is developed based on the principle of mathematical statistics. When using a single-day image pair for analysis, the recognition result of the invariant pixel will be affected by cloud cover, water vapor aerosol content, and extreme weather conditions (such as heavy rain and snow, sand and dust, etc.) Therefore, the correlation coefficient, residual error and other regression quality evaluation indicators of the regression results of each channel can be used to eliminate the results of data that do not meet the expected effect in the long-term series. For some data pairs, when only some of the channels are disturbed, it is considered that the method eliminates the correlation within the channel of the same phase sensor and the correlation between different channels of different sensors, and the calibration and analysis of the channels are independent. So that such data pairs are preserved for valid channel data results. The long-term series results show that the method of the present invention can be implemented with nearly continuous automatic cross-calibration.

The satellite remote sensor cross calibration device based on invariant pixels provided by the present invention is described below. The satellite remote sensor cross calibration device based on invariant pixels described below is the same as the satellite remote sensing based on invariant pixels described above The cross-calibration methods of the device can be referred to each other.

As shown in Figure 6, the embodiment of the present invention discloses a satellite remote sensor cross-calibration device based on invariant pixels, including:

The acquisition module 601 is used to determine a sequence of image pairs, each image pair includes multispectral images obtained by observing the same observation scene separately on the same day and at the same time by the remote sensor to be calibrated and the reference remote sensor;

Invariant pixel detection module 602, used to obtain the apparent reflectance of a single pixel in each image pair and input the invariant pixel detection model to obtain the invariant pixel in each image pair;

The spectral matching module 603 is used to obtain the spectral matching factors of the remote sensor to be calibrated and the reference remote sensor, and perform spectral analysis on the reference remote sensor apparent reflectance of the constant pixel in each image pair based on the spectral matching factor matching, determining the spectrally corrected apparent reflectance of the reference remote sensor for the invariant pixel in each image pair;

The regression module 604 is used to perform orthogonal regression on the spectrally corrected apparent reflectance of the reference remote sensor of the constant pixel in each image pair and the corresponding apparent reflectance of the remote sensor to be calibrated, and determine the cross calibration scalar coefficient.

The satellite remote sensor cross-calibration device based on invariant pixels in the embodiment of the present invention aims at the calibration problems of optical remote sensing imagers and the recalibration requirements of historical data, and proposes a remote sensor cross-calibration device based on intelligent selection of invariant pixels. label method. This method uses the simultaneous satellite scene images acquired by two remote sensors. After data screening and preprocessing, the invariant pixel detection model automatically detects the invariant pixel targets in the scene, and calculates the spectral matching factor to correct the gap between the two sensors. The linear fitting relationship of the apparent reflectance of the constant pixel sample is established, and the cross calibration result and its long-term series calibration trend are obtained.

It can be understood that the image pair sequence determined in the acquisition module 601 includes:

Collect the multi-spectral images obtained from the observations of the same observation scene by the remote sensor to be calibrated and the reference remote sensor at the same time on the same day when the satellite passes through the border at the same time;

The multispectral image collected by the remote sensor to be calibrated and the multispectral image collected by the reference remote sensor are preprocessed to obtain the target image pair.

Based on the target image pair acquired over multiple days, an image pair sequence is determined.

Understandably, preprocessing includes resolution resampling, rasterization, and removal of invalid cells.

Understandably, the removal of invalid cells includes:

Eliminate cloud pollution pixels, water body target pixels, and pixels with satellite observation zenith angles greater than or equal to 30°.

It can be understood that the invariant pixel detection model is established based on the iterative weighted multivariate change detection IR-MAD method.

It can be understood that the invariant pixel detection module 602 includes:

Obtain the detection results of a single pixel in each image pair and calculate the apparent reflectance of the pixel;

Input the pixel apparent reflectance of a single pixel in each image pair into the invariant pixel detection model, construct the MAD variable and construct the observation value of a single pixel in each image pair based on the MAD variable;

Determines the invariant cell in each image pair based on the observations for individual cells in each image pair and an invariant probability decision threshold.

It can be understood that the spectral matching factors of the remote sensor to be calibrated and the reference remote sensor obtained in the spectral matching module 603 include:

Obtain the entrance pupil radiance of the remote sensor to be calibrated and the entrance pupil radiance of the reference remote sensor;

Based on the entrance pupil radiance of the remote sensor to be calibrated and the pupil radiance of the reference remote sensor, the spectral matching factor of the remote sensor to be calibrated and the reference remote sensor is determined.

It can be understood that the regression module 604 includes:

Establish a linear fitting relationship between the spectrally corrected apparent reflectance of the reference remote sensor of the constant pixel of each image pair and the apparent reflectance of the remote sensor to be calibrated of the corresponding image pair of constant pixel to perform orthogonal regression;

Based on the slope and intercept of the orthogonal regression, the cross-calibration coefficients obtained were determined.

It can be understood that the regression module 604 also includes:

Carry out long-term cross-calibration for each channel of the long-time series data of the remote sensor to be calibrated;

Based on the results of the cross-calibration of the long-time series, a cross-calibration coefficient of the long-time series is determined.

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 them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: it can still be Modifications are made to the technical solutions described in the foregoing embodiments, or equivalent replacements are made to some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims (10)

1. A satellite remote sensor cross calibration method based on invariant pixels is characterized by comprising the following steps:

determining a sequence of image pairs, wherein each image pair comprises multispectral images obtained by respectively observing the same observation scene by a remote sensor to be calibrated and a reference remote sensor at the same time on the same day;

obtaining the apparent reflectivity of a single pixel in each image pair and inputting the apparent reflectivity into an invariant pixel detection model to obtain an invariant pixel in each image pair;

acquiring spectrum matching factors of a remote sensor to be calibrated and a reference remote sensor, performing spectrum matching on the apparent reflectivity of the reference remote sensor of an invariant pixel in each image pair based on the spectrum matching factors, and determining the spectrum correction apparent reflectivity of the reference remote sensor of the invariant pixel in each image pair;

and performing orthogonal regression on the spectrum correction apparent reflectivity of the reference remote sensor of the invariant pixel element in each image pair and the apparent reflectivity of the corresponding remote sensor to be calibrated to determine a cross calibration coefficient.

2. The invariant pixel-based satellite remote sensor cross-calibration method of claim 1, wherein the determining a sequence of image pairs comprises:

acquiring multispectral images obtained by respectively observing the same observation scene by a remote sensor to be identified and a reference remote sensor on the same day at the same time when the satellite passes the border at the same time;

preprocessing the multispectral image acquired by the remote sensor to be calibrated and the multispectral image acquired by the reference remote sensor to obtain a target image pair;

determining the sequence of image pairs based on the target image pairs acquired over multiple days.

3. The method of claim 2, wherein the preprocessing comprises resolution resampling, rasterization, and culling of invalid pixels.

4. The satellite remote sensor cross-calibration method based on invariant pixels of claim 3, wherein said eliminating invalid pixels comprises:

and eliminating cloud pollution pixels, water body target pixels and pixels with satellite observation zenith angles larger than or equal to 30 degrees.

5. The satellite remote sensor cross-calibration method based on invariant pixel elements of claim 1, wherein said invariant pixel element detection model is established based on an iterative weighted multivariate change detection IR-MAD method.

6. The satellite remote sensor cross-calibration method based on invariant image elements of claim 5, wherein said obtaining the apparent reflectivity of a single image element in each image pair and inputting the apparent reflectivity into an invariant image element detection model, and outputting the invariant image element in each image pair comprises:

acquiring the detection result of a single pixel in each image pair and calculating the apparent reflectivity of the pixel;

inputting the apparent reflectivity of the pixel of the single pixel in each image pair into an invariant pixel detection model, constructing an MAD variable and constructing an observation value of the single pixel in each image pair based on the MAD variable;

and determining the invariant pixel in each image pair based on the observed value and the invariant probability decision threshold of the single pixel in each image pair.

7. The satellite remote sensor cross calibration method based on the invariant pixel according to claim 1, wherein the obtaining of the spectrum matching factors of the remote sensor to be calibrated and the remote reference sensor comprises:

acquiring the entrance pupil radiance of the remote sensor to be calibrated and the entrance pupil radiance of the reference remote sensor;

and determining the spectrum matching factors of the remote sensor to be calibrated and the reference remote sensor based on the entrance pupil radiance of the remote sensor to be calibrated and the entrance pupil radiance of the reference remote sensor.

8. The satellite remote sensor cross-calibration method based on invariant pixel according to claim 1, wherein said performing orthogonal regression on the spectrum corrected apparent reflectivity of the reference remote sensor of the invariant pixel of each image pair and the apparent reflectivity of the remote sensor to be calibrated of the invariant pixel of the corresponding image pair to determine the cross-calibration coefficient comprises:

establishing a linear fitting relationship between the spectrum correction apparent reflectivity of the reference remote sensor of the invariant pixel of each image pair and the apparent reflectivity of the remote sensor to be calibrated of the invariant pixel of the corresponding image pair for orthogonal regression;

and determining to obtain a cross scaling coefficient based on the slope and intercept of the orthogonal regression.

9. The method for cross-scaling a satellite remote sensor based on invariant image elements as claimed in claim 1, wherein said determining a cross-scaling coefficient further comprises:

performing cross calibration of the long-time sequence on each channel of the long-time sequence data of the remote sensor to be calibrated;

and determining the cross scaling coefficient of the long-time sequence based on the cross scaling result of the long-time sequence.

10. The utility model provides a satellite remote sensor cross calibration device based on invariant pixel, its characterized in that includes:

the acquisition module is used for determining an image pair sequence, and each image pair comprises a multispectral image obtained by respectively observing the same observation scene by the remote sensor to be calibrated and the reference remote sensor at the same time on the same day;

the invariant pixel detection module is used for acquiring the apparent reflectivity of a single pixel in each image pair and inputting the apparent reflectivity into the invariant pixel detection model to acquire the invariant pixel in each image pair;

the spectrum matching module is used for acquiring spectrum matching factors of the remote sensor to be calibrated and the reference remote sensor, performing spectrum matching on the apparent reflectivity of the reference remote sensor of the invariant pixel in each image pair based on the spectrum matching factors, and determining the spectrum correction apparent reflectivity of the reference remote sensor of the invariant pixel in each image pair;

and the regression module is used for performing orthogonal regression on the spectrum correction apparent reflectivity of the reference remote sensor of the invariant pixel in each image pair and the apparent reflectivity of the corresponding remote sensor to be calibrated to determine a cross calibration coefficient.

Images