A multi-channel satellite vegetation optical thickness data consistency correction method and system

Abstract

The application discloses a multi-channel satellite vegetation optical thickness data consistency correction method, comprising the following steps: acquiring and preprocessing multi-channel vegetation optical thickness time series data of a source domain sensor and a target domain sensor; based on the preprocessed multi-channel vegetation optical thickness time series data, the effective sample size of each observation channel is counted to adaptively select quantile points, the continuous cumulative distribution functions of the source domain and the target domain are respectively constructed through quantile interpolation and tail extrapolation, and then the cumulative probability data of the source domain and the target domain are obtained; the cumulative probability data is mapped to a Gaussian space, the correlation matrix of the source domain and the target domain is estimated and normalized, and a connection structure transfer matrix is constructed by matrix decomposition; based on the continuous cumulative distribution function and the connection structure transfer matrix, the source domain cumulative distribution mapping, the joint dependence structure correction and the target domain inverse cumulative distribution mapping are performed on the source domain multi-channel vegetation optical thickness time series data to be corrected, and the corrected source domain multi-channel vegetation optical thickness time series data are obtained.

Description

Technical Field

[0001] This invention relates to the fields of satellite remote sensing data processing, spatiotemporal data fusion, and cross-sensor calibration technology, specifically to a method and system for consistency correction of multi-channel satellite vegetation optical thickness data. Background Technology

[0002] Traditional optical vegetation indices can reflect vegetation greenness or photosynthetic capacity, but they are prone to saturation in areas with high biomass and dense vegetation. Furthermore, they are susceptible to the influence of cloud cover, atmospheric conditions, and sunlight, making continuous monitoring of moisture and structural changes within the vegetation canopy difficult. In contrast, vegetation optical depth (VOD) is more sensitive to changes in canopy water content, vegetation structure, and aboveground biomass. Because microwaves have a certain cloud penetration capability and passive microwave observation is independent of sunlight, surface information can be acquired under diurnal and cloudy conditions, providing an important data foundation for large-scale, long-term, and high temporal resolution monitoring of vegetation moisture status and biomass dynamics.

[0003] Currently, researchers have developed various global VOD products based on different spaceborne microwave sensors and inversion algorithms, forming a vegetation microwave remote sensing data system covering different frequencies, spatial resolutions, and observation periods. Among them, the Advanced Microwave Scanning Radiometer (AMSRE) and its successor sensor AMSR2 provide global multi-frequency microwave observation data over a long time span, providing crucial conditions for constructing long-term VOD products since 2002. Although AMSRE and AMSR2 share some continuity in sensor design, scanning methods, and band configurations, differences remain in radiometric calibration, frequency setting, observation noise, and data processing workflows. Therefore, VOD data from AMSRE and AMSR2 may exhibit systematic intergenerational biases, affecting the identification of long-term vegetation change trends and extreme ecological events.

[0004] Currently, correction methods for intergenerational differences in multi-source remote sensing data mainly include linear regression, mean-variance adjustment, quantile mapping, and cumulative distribution function mapping. For multi-channel VOD data from AMSRE and AMSR2, VOD variables under different observation frequencies, orbital conditions, or other observation configurations are not independent but jointly reflect the moisture and structural state of the vegetation canopy in the same pixel. Therefore, consistency correction of multi-source VOD data not only requires adjusting the numerical distribution of each channel but also ensuring that the corrected multi-channel joint dependency structure is consistent with that of the target domain sensor. Summary of the Invention

[0005] To overcome the shortcomings of traditional multi-source satellite vegetation optical thickness data correction, which typically targets a single channel and struggles to simultaneously eliminate distributional biases between different sensors and maintain multi-channel joint dependencies, thus affecting the consistency of long-term data series, this invention provides a multi-channel satellite vegetation optical thickness data consistency correction method and system. It utilizes adaptive empirical distribution construction and Copula dependency structure transfer to describe the joint dependencies between multiple VOD channels, transferring the correlation structure of source domain sensors to target domain sensors, thereby achieving correction of the multi-channel joint correlation structure.

[0006] According to one aspect of the present invention, a method for consistency correction of multi-channel satellite vegetation optical thickness data is provided, comprising: acquiring and preprocessing multi-channel vegetation optical thickness time-series data from a source domain sensor and a target domain sensor; based on the preprocessed multi-channel vegetation optical thickness time-series data, statistically analyzing the effective sample size of each observation channel to adaptively select quantiles, and constructing continuous cumulative distribution functions for the source domain and the target domain respectively through quantile interpolation and tail extrapolation, thereby obtaining cumulative probability data for the source domain and the target domain; mapping the cumulative probability data to Gaussian space, estimating and positively definiteening the correlation matrix between the source domain and the target domain, and constructing a connection structure migration matrix using matrix decomposition; based on the continuous cumulative distribution functions of the source domain and the target domain and the connection structure migration matrix, sequentially performing source domain cumulative distribution mapping, joint dependency structure correction, and target domain inverse cumulative distribution mapping on the source domain multi-channel vegetation optical thickness time-series data to be corrected, thereby obtaining corrected source domain multi-channel vegetation optical thickness time-series data.

[0007] Furthermore, the multi-channel vegetation optical thickness time-series data includes several vegetation optical thickness data obtained from source domain sensors and several vegetation optical thickness data obtained from target domain sensors; wherein, each vegetation optical thickness data is divided into several spatial pixels according to geographic space, each spatial pixel has multiple observation channels, and each observation channel stores the vegetation optical thickness value obtained from the corresponding channel inversion.

[0008] Further, the preprocessing includes: setting the spatial pixels corresponding to vegetation optical thickness values ​​less than or equal to a preset physical effective threshold in the vegetation optical thickness data as invalid values; reading land cover data and setting the spatial pixels corresponding to the permanent snow and ice land cover category in the land cover data as invalid values; wherein, each spatial pixel of the land cover data corresponds one-to-one with each spatial pixel of the vegetation optical thickness data; using spatial pixels and time sub-periods as processing units, the multi-channel vegetation optical thickness time-series data of the source domain sensor and the target domain sensor are divided respectively to obtain the source domain training sample set and the target domain training sample set.

[0009] Furthermore, based on the preprocessed multi-channel vegetation optical thickness time-series data, the effective sample size of each observation channel is statistically analyzed to adaptively select quantiles. Continuous cumulative distribution functions for the source and target domains are constructed through quantile interpolation and tail extrapolation, respectively, thereby obtaining the cumulative probability data for the source and target domains. This includes: for each spatial pixel and each time sub-period, the effective sample size of each observation channel in the corresponding source and target domain training sample sets is statistically analyzed to determine the minimum effective sample size, and a corresponding quantile set is selected from multiple candidate quantile set sets based on the minimum effective sample size; based on the selected quantile set, the quantile values ​​of the source and target domain training sample sets in each observation channel are calculated respectively; through interpolation between adjacent quantile values ​​and tail fitting outside the boundary quantile values, the source domain empirical cumulative distribution function and the target domain empirical cumulative distribution function are established respectively; using the source and target domain empirical cumulative distribution functions, the source and target domain training samples of each observation channel are transformed into probability spaces to obtain source domain probability samples and target domain probability samples.

[0010] Furthermore, the interpolation between adjacent quantiles and the tail fitting outside the boundary quantiles include: determining the cumulative probability for training samples between the second lowest quantile and the second highest quantile using linear interpolation; for training samples below the second lowest quantile or above the second highest quantile, performing a linear function fitting using the ranking value of the corresponding tail training samples and the tail cumulative probability; and when the tail training sample size is less than the tail threshold, using the boundary probability corresponding to the tail threshold instead of the tail cumulative probability to perform a linear function fitting.

[0011] Further, the cumulative probability data is mapped to Gaussian space, the source domain and target domain correlation matrices are estimated and positively definite, and a connection structure transfer matrix is ​​constructed using matrix decomposition. This includes: selecting joint effective samples from the source domain and joint effective samples from the target domain from all source domain probability samples and target domain probability samples, respectively, where the vegetation optical thickness values ​​obtained from each observation channel under the same spatial pixel are all effective values; using the inverse function of the standard normal cumulative distribution function, transforming the joint effective samples from the source domain and joint effective samples from the target domain to Gaussian space to obtain source domain Gaussian samples and target domain Gaussian samples; determining the source domain channel correlation matrix and the target domain channel correlation matrix based on the source domain Gaussian samples and the target domain Gaussian samples, respectively, and performing positive definite processing on the two channel correlation matrices; performing matrix decomposition on the positively definite source domain channel correlation matrix and the target domain channel correlation matrix to obtain the source domain decomposition matrix and the target domain decomposition matrix; and constructing a connection structure transfer matrix that transfers the source domain channel correlation structure to the target domain channel correlation structure based on the source domain decomposition matrix and the target domain decomposition matrix.

[0012] Furthermore, based on the source domain decomposition matrix and the target domain decomposition matrix, a connection structure transfer matrix is ​​constructed to transfer the source domain channel-related structure to the target domain channel-related structure. The corresponding formula is:

[0013] A=C 目标 C 源 -1 ,

[0014] Where A is the connection structure transition matrix; C 源 -1 It is the inverse matrix of the source domain decomposition matrix; C 目标 It is the target domain decomposition matrix.

[0015] Furthermore, the method further includes: calculating the daily average value of the source domain multi-channel vegetation optical thickness time series data before correction and the daily average value of the source domain multi-channel vegetation optical thickness time series data after correction; comparing the daily average value of the source domain multi-channel vegetation optical thickness time series data before and after correction with the daily average value of the target domain multi-channel vegetation optical thickness time series data to verify the corrected source domain multi-channel vegetation optical thickness time series data.

[0016] According to one aspect of the present invention, a multi-channel satellite vegetation optical thickness data consistency correction system is provided, comprising: a data acquisition and preprocessing module for acquiring and preprocessing multi-channel vegetation optical thickness time-series data from a source domain sensor and a target domain sensor; an adaptive empirical distribution construction module for, based on the preprocessed multi-channel vegetation optical thickness time-series data, statistically analyzing the effective sample size of each observation channel to adaptively select quantiles, and constructing continuous cumulative distribution functions for the source domain and the target domain respectively through quantile interpolation and tail extrapolation, thereby obtaining cumulative probability data for the source domain and the target domain; a dependency structure transfer module for mapping the cumulative probability data to Gaussian space, estimating and positively determining the correlation matrix between the source domain and the target domain, and constructing a connection structure transfer matrix using matrix decomposition; and an inverse cumulative distribution mapping module for, based on the continuous cumulative distribution functions of the source domain and the target domain and the connection structure transfer matrix, sequentially performing source domain cumulative distribution mapping, joint dependency structure correction, and target domain inverse cumulative distribution mapping on the source domain multi-channel vegetation optical thickness time-series data to be corrected, thereby obtaining corrected source domain multi-channel vegetation optical thickness time-series data.

[0017] According to one aspect of the present invention, a non-transitory computer-readable storage medium is provided, the non-transitory computer-readable storage medium storing computer instructions that cause the computer to execute the aforementioned multi-channel satellite vegetation optical thickness data consistency correction method.

[0018] The present invention discloses a method for consistency correction of multi-channel satellite vegetation optical thickness data. First, multi-channel time-series vegetation optical thickness data from source and target domain sensors are acquired and preprocessed. Then, using pixels and months as processing units, quantiles are adaptively selected based on the effective sample size. Continuous cumulative distribution functions for the source and target domains are constructed through quantile interpolation and tail extrapolation, respectively. The multi-channel data is mapped to Gaussian space, and the correlation matrices of the source and target domains are estimated and positively definite. A Copula structure migration matrix is ​​constructed using matrix decomposition. Finally, source domain cumulative distribution mapping, joint dependency structure correction, and target domain inverse cumulative distribution mapping are performed sequentially to obtain the corrected vegetation optical thickness data. The corrected data can correct the deviation between the source and target domain sensors. This invention can simultaneously correct single-channel edge distribution and multi-channel correlation structure, providing a consistent data foundation for the construction of long-term cross-sensor vegetation optical thickness series and dynamic monitoring of vegetation.

[0019] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0020] 1. Improve the consistency of long-term VOD series across sensors.

[0021] Existing AMSRE and AMSR2 vegetation optical thickness data differ in radiometric calibration, frequency settings, observation noise, and inversion processes, potentially leading to unnatural abrupt changes at sensor replacement points. This invention uses AMSR2 as the target domain to perform statistical distribution and joint dependency structure correction on AMSRE data. This reduces systematic bias caused by sensor differences, improves the continuity and consistency of long-term VOD data, and provides a more reliable data foundation for research on long-term changes in vegetation moisture, aboveground biomass, and ecosystems.

[0022] 2. Simultaneously correct single-channel edge distribution and multi-channel correlation structure.

[0023] Traditional quantile mapping or cumulative distribution function correction methods typically process each observed variable independently. While these methods can improve the mean, variance, and quantile distributions of individual channels, they may alter the original correlations between different observed channels. This invention first uses a continuous cumulative distribution function to characterize the marginal distributions of each VOD variable (i.e., the numerical distribution of a single VOD variable itself). Then, it establishes a multi-channel joint dependency structure between AMSRE and AMSR2 using the Gaussian Copula method, and transfers the target domain correlation structure to the source domain data. Therefore, this invention not only makes the numerical distributions of each channel after correction approximate AMSR2, but also ensures that the correlations between multiple channels remain consistent with AMSR2.

[0024] 3. Improve the stability of distribution modeling under different sample conditions.

[0025] The number of effective samples varies significantly across different pixels and months due to factors such as missing data, land surface type, and observation conditions. Using a fixed number of quantiles uniformly may lead to duplicate quantile values, unstable interpolation, or overfitting when samples are insufficient. This invention adaptively selects a set of quantiles with different densities based on the minimum effective sample size of the source and target domains: more quantiles are used when samples are sufficient to describe distribution details, while fewer quantiles are used when samples are limited to improve estimation stability. This method can adapt to sample conditions of different pixels and months, improving the reliability of cumulative distribution function construction and data correction.

[0026] 4. Improve the mapping effect of distribution tails and extreme VOD values.

[0027] Conventional quantile mapping methods typically rely on interpolation within the quantile range, which has limited ability to handle high and low value samples outside the internal interval, easily leading to extreme value accumulation, mapping interruption, or unreasonable extrapolation. This invention employs quantile linear interpolation for the middle of the distribution, and uses corresponding training samples for fitting and extrapolation at the low and high value tails respectively, limiting the correction results to a reasonable range of values ​​within the target domain training samples. Therefore, this invention can reduce the impact of abnormal extrapolated values ​​on the correction results while maintaining mapping continuity. Attached Figure Description

[0028] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0029] Figure 1 A flowchart for vegetation optical thickness consistency correction provided in an embodiment of the present invention;

[0030] Figure 2 A schematic diagram of the daily average curves of VOD data before and after correction provided in an embodiment of the present invention;

[0031] Figure 3 The scatter plot of VOD data before and after correction is provided in an embodiment of the present invention. Detailed Implementation

[0032] It should be noted that:

[0033] The terms “comprising” and “having”, and any variations thereof, in the specification, claims, and drawings of this invention are intended to cover a non-exclusive inclusion, such as a process, method, system, product, or apparatus that includes a series of steps or units, not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0034] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices. The flowcharts shown in the accompanying drawings are merely illustrative and do not necessarily include all content and operations / steps, nor do they necessarily have to be performed in the described order. For example, some operations / steps can be decomposed, while others can be combined or partially combined; therefore, the actual execution order may change depending on the specific circumstances.

[0035] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. In addition, the technical features of the various embodiments or individual embodiments provided by the present invention can be arbitrarily combined to form new technical solutions. Such combinations are not bound by the order of steps and / or structural composition patterns, but must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0036] Please refer to the appendix. Figure 1 This invention provides a method for consistency correction of multi-channel satellite vegetation optical thickness data, which sequentially performs four main steps: data collection, adaptive empirical distribution construction, Copula-dependent structure transfer, and inverse cumulative distribution mapping. Specifically:

[0037] Step S1: Acquire and preprocess multi-channel vegetation optical thickness time-series data from the source domain sensor and the target domain sensor.

[0038] Step S1 aims to collect and preprocess data. Both the source domain sensor and the target domain sensor are satellite sensors mounted on satellites. The multi-channel vegetation optical thickness time-series data includes several vegetation optical thickness data points obtained from the source domain sensor and several vegetation optical thickness data points obtained from the target domain sensor (i.e., VOD data). The VOD data is raster data, which is divided into a large number of regular spatial pixels according to geographic space (0.25-degree resolution in this embodiment). Each spatial pixel has multiple observation channels (four observation channels in this embodiment), and each observation channel stores the vegetation optical thickness value obtained from the corresponding channel inversion. The source domain refers to the VOD data that needs to be corrected (AMSRE VOD data); the target domain refers to the target data to be corrected, that is, the level of data to be corrected (AMSR2 VOD data).

[0039] Preprocessing includes: setting spatial pixels corresponding to vegetation optical thickness values ​​less than or equal to a preset physical effective threshold in the vegetation optical thickness data to invalid values; reading land cover data and setting spatial pixels corresponding to the permanent snow and ice land cover category in the land cover data to invalid values; using spatial pixels and time sub-periods as processing units, dividing the multi-channel vegetation optical thickness time-series data of the source domain sensor and the target domain sensor respectively to obtain the source domain training sample set and the target domain training sample set. In this embodiment, the preset physical effective threshold is 0, and the time sub-period is months.

[0040] It should be noted that land cover data is also raster data. Each spatial cell in the land cover data stores a category code, which represents the land cover type (woodland, grassland, bare land, permanent snow and ice, water body, etc.) corresponding to that geographic location. The land cover raster itself does not store vegetation optical thickness; it is used to identify the spatial location of permanent snow and ice, enabling masking filtering of temporal data on vegetation optical thickness. It should also be noted that land cover data is external auxiliary data independent of VOD data. After georegistration, each spatial cell in the land cover data achieves a one-to-one geographic correspondence with each spatial cell in the VOD data.

[0041] This embodiment uses a global land area as the case study scope, and the VOD data obtained from the AMSRE and AMSR2 satellite sensors as the processing objects, implemented using MATLAB programming. AMSRE serves as the source domain sensor, with data spanning from 2002 to 2011, while AMSR2 serves as the target domain sensor, with data spanning from 2012 to 2022. The vegetation optical thickness data from both satellite sensors include four VOD observation variables (the four observation variables refer to the VOD data obtained from the four observation channels, and the observation channels refer to different orbits and frequencies) to characterize the microwave attenuation characteristics of the vegetation canopy under different observation times or conditions. AMSRE data serves as the source domain data to be corrected, and AMSR2 data serves as the target domain data providing the target statistical distribution and multi-channel dependency structure. Simultaneously, land cover data is acquired. This implementation uses the 2022 MCD12Q2 International Geosphere-Biosphere Programme (IGBP) classification data (https: / / lpdaac.usgs.gov / products / mcd12c1v061 / ), and sets the spatial pixels corresponding to the permanent snow and ice land cover category to invalid values.

[0042] Step S2: Based on the preprocessed multi-channel vegetation optical thickness time series data, the effective sample size of each observation channel is statistically analyzed to adaptively select quantiles. Continuous cumulative distribution functions of the source domain and target domain are constructed by quantile interpolation and tail extrapolation, respectively, thereby obtaining the cumulative probability data of the source domain and target domain.

[0043] Step S2 aims to construct an adaptive empirical distribution. Understandably, each source domain training sample in the source domain training sample set is a slice of source domain VOD data under the same spatial pixel, and each target domain training sample in the target domain training sample set is a slice of target domain VOD data under the same spatial pixel. Since each spatial pixel has multiple observation channels, correspondingly, each source / target domain training sample also has multiple observation channels, and each observation channel stores the vegetation optical thickness value retrieved from the corresponding channel. The rule for determining whether a source / target domain training sample is valid under any observation channel is: the vegetation optical thickness value of the source / target domain training sample under that observation channel is greater than zero. The effective sample size refers to the number of valid source / target domain training samples under each observation channel, and each source / target domain training sample can be valid under multiple observation channels simultaneously.

[0044] Step S21: For each spatial pixel and each time sub-period, the effective sample size of each observation channel in the corresponding source domain training sample set and target domain training sample set is counted to determine the minimum effective sample size, and the corresponding quantile set is selected from multiple candidate quantile sets according to the minimum effective sample size.

[0045] Step S21 aims to perform effective sample statistics and adaptive quantile selection. The effective sample count refers to the number of effective samples for a single observation channel within a spatial pixel and a time sub-period. Each observation channel within a spatial pixel and a time sub-period corresponds to its own effective sample count, and the minimum effective sample count among all observation channels is the minimum effective sample size. Each spatial pixel and each time sub-period corresponds to a minimum effective sample size, and a corresponding quantile set is selected from multiple candidate quantile sets based on this minimum effective sample size for each spatial pixel and each time sub-period. Different spatial pixels and time sub-periods correspond to different quantile sets. The quantile set is used to calculate the quantile values ​​of the VOD data retrieved from each observation channel within the corresponding spatial pixel and time sub-period. It is understandable that the quantile set used for each observation channel within the same spatial pixel and the same time sub-period is the same, but the VOD data retrieved from each observation channel is different; therefore, the quantile values ​​of the VOD data retrieved from each observation channel are also different.

[0046] This embodiment uses spatial pixels and months as processing units, respectively counting the number of effective samples in the source and target domains across four observation channels, and using the minimum effective sample size (i.e., the minimum value among the effective sample sizes) as the quantile selection criterion. This invention pre-defines multiple sample size intervals and corresponding candidate quantile sets. When the minimum effective sample size falls within different sample size intervals, candidate quantile sets with different quantile densities are selected, where a larger minimum effective sample size results in a selection set containing more quantiles. In this embodiment, 13 quantiles are used when the minimum effective sample size is ≥200; 11 quantiles are used when 100 ≤ minimum effective sample size <200; 6 quantiles are used when 50 ≤ minimum effective sample size <100; and 5 quantiles are used when 20 ≤ minimum effective sample size <50 (specifically as shown in Table 1). Understandably, more quantiles are used when the sample size is sufficient to describe distribution details, while fewer quantiles are used when the sample size is small to reduce duplicate quantile values ​​and the risk of overfitting.

[0047] Table 1. Correspondence between sample size and quantiles

[0048]

[0049] Step S22: Based on the selected set of quantiles, calculate the quantile values ​​of the source domain training sample set and the target domain training sample set on each observation channel. Through interpolation between adjacent quantile values ​​and tail fitting outside the boundary quantile values, establish the source domain empirical cumulative distribution function and the target domain empirical cumulative distribution function respectively.

[0050] Step S22 aims to calculate quantile values ​​and construct the continuous cumulative distribution function. Specifically, based on the quantiles and corresponding cumulative probabilities of each quantile in the selected set of quantiles, the quantile values ​​of each VOD variable (i.e., the VOD data obtained from each observation channel) in the source and target domains are calculated respectively. Interpolation between adjacent quantile values ​​refers to linear interpolation between quantile values ​​and probability nodes for samples within the distribution; tail fitting outside the boundary quantiles refers to linear function fitting and extrapolation using the corresponding tail training samples for low-value tails and high-value tails, respectively.

[0051] ,

[0052] Among them, F j (X) represents the cumulative distribution function of the j-th VOD variable; X represents the vegetation optical thickness value (i.e., VOD value) to be converted; q j,l q represents the l-th quantile value; j,l+1 Represents l+1 quantile values; p l Indicates with q j,l The corresponding cumulative probability; p l+1 Indicates with q j,l+1 The corresponding cumulative probability.

[0053] Further, the interpolation between adjacent quantiles and the tail fitting outside the boundary quantiles include: determining the cumulative probability using linear interpolation for training samples between the second lowest quantile and the second highest quantile; for training samples below the second lowest quantile or above the second highest quantile, performing a linear function fitting using the ranking value of the corresponding tail training samples and the tail cumulative probability; and when the tail training sample size is less than the tail threshold, using the boundary probability corresponding to the tail threshold instead of the tail cumulative probability for linear function fitting. It should be noted that tail training samples refer to training samples whose quantiles are equal to the first lowest quantile and the first highest quantile, and the training samples are either source domain training samples or target domain training samples. The tail training sample size refers to the number of training samples whose quantiles are equal to the first lowest quantile and the first highest quantile. The tail training samples are sorted according to the size of the VOD value. The ranking value refers to the sorted VOD value, and the cumulative probability represents the relative position of a certain VOD value in all samples. The VOD values ​​must first be sorted from smallest to largest before a cumulative probability corresponding to their size can be assigned to each VOD value. Wherein, the first lowest quantile is less than the second lowest quantile, the second lowest quantile is less than the second highest quantile, and the second highest quantile is less than the first highest quantile. The tail threshold represents the minimum number of samples required for tail fitting. The first lowest quantile, first highest quantile, second lowest quantile, second highest quantile, and tail threshold can all be set according to actual needs and are not limited here.

[0054] Step S23: Using the source domain empirical cumulative distribution function and the target domain empirical cumulative distribution function, the source domain training samples and target domain training samples of each observation channel are transformed into the probability space to obtain source domain probability samples and target domain probability samples (i.e., cumulative probability data of the source domain and the target domain).

[0055] Step S23 aims to transform the data to a uniform probability space. Specifically, using the source domain empirical cumulative distribution function and the target domain empirical cumulative distribution function, the vegetation optical thickness values ​​(i.e., VOD values) corresponding to the source domain training samples and target domain training samples under each observation channel are converted into cumulative probabilities in the range of 0-1. To avoid the standard normal inverse function producing infinite values ​​when the probability is equal to 0 or 1, the probability is limited to […]. , Within the scope of this embodiment, .

[0056] U 源 =F 源 (X 源 ),

[0057] U 目标 =F 目标 (X 目标 ),

[0058] Among them, U 源 U represents the cumulative probability after the AMSRE transformation; 目标 X represents the cumulative probability after AMSR2 transformation; 源 Indicates the raw VOD value of AMSRE; X 目标 Indicates the raw VOD value of AMSR2; F 源 F represents the source domain empirical cumulative distribution function of AMSRE; 目标 This represents the empirical cumulative distribution function of the target domain of AMSR2.

[0059] In this embodiment, the observation date is divided into 12 month groups based on the natural month to which it belongs. The period from 2003 to 2010 is used as the AMSRE source domain training period, and the period from 2013 to 2020 is used as the AMSR2 target domain training period. For each spatial pixel and each month, source domain training samples and target domain training samples for four VOD variables are extracted respectively. The number of effective samples for the four VOD variables of AMSRE and AMSR2 within the current pixel and month is counted, and the minimum number of effective samples among all VOD variables is used as the quantile selection criterion. Based on the selected quantiles, the quantile values ​​of each VOD variable of AMSRE and AMSR2 are calculated, and the empirical cumulative distribution function of each VOD variable is constructed. For VOD values ​​in the middle of the distribution, the cumulative probability is calculated by linear interpolation between adjacent quantile values; for samples below the lower quantile value or above the upper quantile value, linear fitting and extrapolation are performed using training samples from the lower and upper tails respectively. Thus, the original VOD values ​​of each variable of AMSRE and AMSR2 are converted into cumulative probability values ​​in the range of 0-1.

[0060] Step S3: Map the cumulative probability data to Gaussian space, estimate and positive definite the correlation matrix between the source domain and the target domain, and construct the connection structure transfer matrix (i.e., the Copula structure transfer matrix) using matrix decomposition.

[0061] Step S3 aims to perform Copula dependency structure migration. Specifically, it includes the following steps:

[0062] Step S31: From all source domain probability samples and target domain probability samples, select source domain joint effective samples and target domain joint effective samples where the vegetation optical thickness values ​​obtained by each observation channel under the same spatial pixel are all effective values; use the inverse function of the standard normal cumulative distribution function to transform the source domain joint effective samples and target domain joint effective samples to Gaussian space to obtain source domain Gaussian samples and target domain Gaussian samples.

[0063] Step S31 aims to perform joint effective sample screening and Gaussian space transformation. In this embodiment, only source domain probability samples and target domain probability samples whose vegetation optical thickness values ​​obtained from the inversion of the four observation channels under the same spatial pixel are all valid values ​​(i.e., VOD values ​​are not less than 0, indicating no permanent snow or ice) are retained; these are the source domain joint effective samples and target domain joint effective samples. When the number of source domain joint effective samples or target domain joint effective samples is less than 20, the correction of the current spatial pixel and month is stopped. Otherwise, the source domain probability samples and target domain probability samples are transformed to Gaussian space using the inverse function of the standard normal cumulative distribution function.

[0064] Z 源 = (U 源 ),

[0065] Z 目标 = (U 目标 ),

[0066] Among them, Z 源 and Z 目标 These represent the representations of AMSRE and AMSR2 data in Gaussian space, namely, Gaussian samples in the source domain and Gaussian samples in the target domain, respectively. The inverse function of the cumulative distribution function of the standard normal distribution, also known as the standard normal quantile function; U 源 and U 目标 These represent the source domain probability sample and the target domain probability sample, respectively.

[0067] Step S32: Determine the source domain channel correlation matrix and the target domain channel correlation matrix based on the source domain Gaussian sample and the target domain Gaussian sample, respectively, and perform positive definite processing on the two channel correlation matrices.

[0068] Step S32 aims to perform correlation matrix estimation and positive definiteness. In this embodiment, based on the joint samples in Gaussian space (i.e., source domain Gaussian samples and target domain Gaussian samples), the channel correlation matrices between the four observation channels in the source and target domains are calculated respectively, resulting in the source domain channel correlation matrix and the target domain channel correlation matrix. The two channel correlation matrices are then subjected to positive definiteness processing, including: performing non-finite value processing and symmetry, setting the diagonal elements to 1, and adding a regularization term λ to improve the numerical stability of the matrix decomposition. In this embodiment, λ = .

[0069] R 源 =Corr(Z 源 ),

[0070] R 目标 =Corr(Z 目标 ),

[0071] Among them, R 源 and R 目标 represents the correlation matrix of the four VOD variables AMSRE and AMSR2 in Gaussian space; Corr represents the correlation calculation.

[0072] Step S33: Perform matrix decomposition on the positively definite source domain channel correlation matrix and target domain channel correlation matrix respectively to obtain the source domain decomposition matrix and the target domain decomposition matrix; based on the source domain decomposition matrix and the target domain decomposition matrix, construct the connection structure transfer matrix that transfers the source domain channel correlation structure to the target domain channel correlation structure.

[0073] Step S33 aims to perform Cholesky decomposition (i.e., Cholesky decomposition or square root method) and construct the structure transfer matrix. The source domain channel correlation structure refers to the joint structure of the source domain multi-channel vegetation optical thickness data to be corrected in Gaussian space, while the target domain channel correlation structure refers to the joint structure of the target domain multi-channel vegetation optical thickness data in Gaussian space. The joint structure refers to the joint dependency between multiple observation channels, i.e., how other observation channels change in tandem when one observation channel changes. This joint dependency is usually represented by a correlation coefficient matrix in Gaussian space. This Copula structure transfer matrix is ​​used to transform the joint structure of the AMSRE data to be corrected in Gaussian space.

[0074] In this embodiment, the positive-definite source domain channel correlation matrix and the target domain channel correlation matrix are subjected to lower triangular Cholesky decomposition, respectively, to obtain two lower triangular decomposition matrices. Based on these two lower triangular decomposition matrices, a Copula structure transition matrix is ​​constructed to transform the AMSRE joint correlation structure into the AMSR2 joint correlation structure:

[0075] A=C 目标 C 源 -1 ,

[0076] Where A is the Copula structure transition matrix (i.e., the Copula structure transition matrix) that transforms the AMSRE joint correlation structure into the AMSR2 joint correlation structure; C 源 -1 It is the inverse matrix of AMSRE after Cholesky decomposition (i.e., the inverse matrix of the source domain decomposition matrix); C 目标 It is the matrix obtained by Cholesky decomposition of AMSR2 (i.e., the target domain decomposition matrix).

[0077] Step S4: Based on the continuous cumulative distribution function of the source domain and the target domain and the connection structure migration matrix, the source domain multi-channel vegetation optical thickness time series data to be corrected is sequentially subjected to source domain cumulative distribution mapping, joint dependency structure correction and target domain inverse cumulative distribution mapping to obtain the corrected source domain multi-channel vegetation optical thickness time series data.

[0078] Step S4 aims to perform source domain cumulative distribution mapping, joint dependency structure correction, and target domain inverse cumulative distribution mapping to correct the bias between the source domain sensor and the target domain sensor. It should be noted that the source domain multi-channel vegetation optical thickness time-series data to be corrected is the same as the multi-channel vegetation optical thickness time-series data obtained from the source domain sensor inversion in step S1. Step S4 specifically includes the following steps:

[0079] Step S41: Using the source domain empirical cumulative distribution function, the source domain multi-channel vegetation optical thickness data to be corrected is transformed into probability space, and then into Gaussian space; the joint dependency structure of each observation channel is corrected using the connection structure migration matrix to obtain the source domain Gaussian variables that have completed the joint dependency structure correction.

[0080] Step S41 aims to apply the correction model to all AMSRE observation data. In this embodiment, for AMSRE data from 2002 to 2011, the VOD to be corrected is first transformed to probability space using the source domain empirical cumulative distribution function, and then to Gaussian space; subsequently, the joint dependency structure of the four observation channels using the Copula structure migration matrix (joint dependency structure refers to the relationship of common change among multiple observation channels, i.e., when the VOD data of one channel changes, the other channels usually change in the same direction and to what extent) is used to obtain the source domain Gaussian variables that complete the joint dependency structure correction:

[0081] Z 校正 =Z 源 A T ,

[0082] Among them, Z 校正 This represents the corrected AMSRE data in Gaussian space; A T This represents the transpose of A.

[0083] Step S42: Using the standard normal cumulative distribution function, transform the source domain Gaussian variables that have undergone joint dependency structure correction back into the uniform probability space to obtain the corrected source domain cumulative probability:

[0084] U 校正 =ΦZ 校正 ,

[0085] Among them, U 校正 Φ represents the cumulative probability after the AMSRE data has been corrected and transformed; Φ represents the standard normal cumulative distribution function.

[0086] Step S43: Using the inverse function of the empirical cumulative distribution function of the target domain, the corrected cumulative probability of the source domain is inversely mapped to the corrected multi-channel vegetation optical thickness time series data of the source domain.

[0087] Step S43 aims to perform inverse cumulative distribution mapping of the target domain. In this embodiment, the inverse cumulative distribution function of each VOD variable in the AMSR2 target domain is used to restore the corrected probability to a VOD value with actual physical meaning. The inverse mapping of the inverse function of the empirical cumulative distribution function of the target domain includes: linear interpolation between adjacent target quantiles is used for the internal probability interval; for the low probability tail and the high probability tail, a linear function extrapolation from the probability to the vegetation optical thickness value is performed using the tail samples of the target domain to obtain the corrected source domain multi-channel vegetation optical thickness time-series data.

[0088] X 校正 =F 目标 -1 U 校正 ,

[0089] Among them, X 校正 Represents the corrected VOD; F 目标 -1 This represents the inverse function of the cumulative distribution function of AMSR2.

[0090] To avoid tail extrapolation producing results that exceed the reasonable range of the target domain sensor, the correction value is limited to the minimum and maximum values ​​of the corresponding target domain training samples, and correction results less than or equal to 0 are reset to null values.

[0091] Through the above steps, this invention corrects the differences in the numerical distribution of individual VOD variables between AMSRE and AMSR2, while maintaining and migrating the joint dependency structure between multiple VOD variables, reducing unnatural time series mutations caused by sensor replacement, and providing technical support for constructing continuous and consistent long-term VOD products.

[0092] Furthermore, in some feasible embodiments, the method further includes: calculating the daily average value of the source domain multi-channel vegetation optical thickness time-series data before correction and the daily average value of the source domain multi-channel vegetation optical thickness time-series data after correction, respectively; comparing the daily average value of the source domain multi-channel vegetation optical thickness time-series data before and after correction with the daily average value in the target domain multi-channel vegetation optical thickness time-series data, in order to verify the corrected source domain multi-channel vegetation optical thickness time-series data. In this embodiment, in order to verify the corrected target domain multi-channel vegetation optical thickness data, the daily average value of the AMSRE VOD data before and after correction of the four variables is calculated, and compared with the daily average value of AMSR2 (e.g., Figure 2As shown in the figure. It can be seen that before correction, the daily average VOD of AMSRE was lower than that of AMSR2, especially for ascending orbits. After correction, the daily average VOD of AMSRE is closer to that of AMSR2. In addition, the correlation (including R, ubRMSE, RMSE, and Bias) between the average VOD of 2010 (representing AMSRE) and the VOD of 2013 (representing AMSR2) before and after correction was calculated and a scatter plot was drawn (i.e.,...). Figure 3 It can be seen that the corrected AMSRE VOD has a stronger correlation with AMSR2 VOD, lower error, and smaller bias.

[0093] Based on the same technical concept as the aforementioned embodiments, the present invention also provides a multi-channel satellite vegetation optical thickness data consistency correction system, comprising: a data acquisition and preprocessing module, used to acquire and preprocess multi-channel vegetation optical thickness time-series data from source domain sensors and target domain sensors; an adaptive empirical distribution construction module, used to, based on the preprocessed multi-channel vegetation optical thickness time-series data, statistically analyze the effective sample size of each observation channel to adaptively select quantiles, and construct continuous cumulative distribution functions for the source domain and target domain respectively through quantile interpolation and tail extrapolation, thereby obtaining cumulative probability data for the source domain and target domain; a dependency structure transfer module, used to map the cumulative probability data to Gaussian space, estimate and positively definiteen the correlation matrix between the source domain and target domain, and construct a connection structure transfer matrix using matrix decomposition; and an inverse cumulative distribution mapping module, used to, based on the continuous cumulative distribution functions of the source domain and target domain and the connection structure transfer matrix, sequentially perform source domain cumulative distribution mapping, joint dependency structure correction, and target domain inverse cumulative distribution mapping on the source domain multi-channel vegetation optical thickness time-series data to be corrected, to obtain corrected source domain multi-channel vegetation optical thickness time-series data.

[0094] Based on the same technical concept as the foregoing embodiments, the present invention also provides a non-transitory computer-readable storage medium storing computer instructions that cause the computer to execute the aforementioned multi-channel satellite vegetation optical thickness data consistency correction method.

[0095] In summary, this invention discloses a method for consistency correction of multi-channel satellite vegetation optical thickness data. Traditional multi-source satellite vegetation optical thickness data correction typically targets a single channel, making it difficult to simultaneously eliminate distribution biases between different sensors and maintain multi-channel joint dependencies, thus affecting the consistency of long-term data series. First, multi-channel vegetation optical thickness time-series data from source and target domain sensors are acquired and preprocessed. Then, using pixels and months as processing units, quantiles are adaptively selected based on the effective sample size. Continuous cumulative distribution functions for the source and target domains are constructed using quantile interpolation and tail extrapolation, respectively. The multi-channel data is mapped to Gaussian space, and the correlation matrices of the source and target domains are estimated and positively definite. A Copula structure migration matrix is ​​constructed using matrix decomposition. Finally, source domain cumulative distribution mapping, joint dependency structure correction, and target domain inverse cumulative distribution mapping are sequentially performed to obtain the corrected vegetation optical thickness data. The corrected data can correct the biases between source and target domain sensors. This invention can simultaneously correct single-channel edge distribution and multi-channel correlation structures, providing a consistent data foundation for the construction of long-term vegetation optical thickness series across sensors and for dynamic vegetation monitoring.

[0096] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the technical solutions of the embodiments of the present invention.

Claims

1. A method for consistency correction of multi-channel satellite vegetation optical thickness data, characterized in that, include: Acquire and preprocess multi-channel vegetation optical thickness time-series data from source and target domain sensors; Based on the preprocessed multi-channel vegetation optical thickness time series data, the effective sample size of each observation channel is statistically analyzed to adaptively select quantiles. Continuous cumulative distribution functions of the source domain and target domain are constructed by quantile interpolation and tail extrapolation, respectively, thereby obtaining the cumulative probability data of the source domain and target domain. The cumulative probability data is mapped to Gaussian space, the correlation matrix between the source domain and the target domain is estimated and positively definite, and the connection structure transition matrix is ​​constructed by matrix decomposition. Based on the continuous cumulative distribution function of the source domain and the target domain and the connection structure transfer matrix, the source domain multi-channel vegetation optical thickness time series data to be corrected are sequentially subjected to source domain cumulative distribution mapping, joint dependency structure correction and target domain inverse cumulative distribution mapping to obtain the corrected source domain multi-channel vegetation optical thickness time series data.

2. The method for consistency correction of multi-channel satellite vegetation optical thickness data as described in claim 1, characterized in that, The multi-channel vegetation optical thickness time series data includes several vegetation optical thickness data obtained from source domain sensors and several vegetation optical thickness data obtained from target domain sensors; each vegetation optical thickness data is divided into several spatial pixels according to geographic space, each spatial pixel has multiple observation channels, and each observation channel stores the vegetation optical thickness value obtained from the corresponding channel inversion.

3. The method for consistency correction of multi-channel satellite vegetation optical thickness data as described in claim 2, characterized in that, The preprocessing includes: Set the spatial pixels corresponding to vegetation optical thickness values ​​that are less than or equal to the preset physical effective threshold in the vegetation optical thickness data to invalid values. Read the land cover data and set the spatial pixels corresponding to the permanent snow and ice land cover category in the land cover data to invalid values; each spatial pixel in the land cover data corresponds one-to-one with each spatial pixel in the vegetation optical thickness data; Using spatial pixels and temporal sub-periods as processing units, the multi-channel vegetation optical thickness time-series data of the source domain sensor and the target domain sensor are divided to obtain the source domain training sample set and the target domain training sample set.

4. The method for consistency correction of multi-channel satellite vegetation optical thickness data as described in claim 2, characterized in that, Based on the preprocessed multi-channel vegetation optical thickness time-series data, the effective sample size of each observation channel was statistically analyzed to adaptively select quantiles. Continuous cumulative distribution functions for the source and target domains were constructed using quantile interpolation and tail extrapolation, respectively, thus obtaining the cumulative probability data for the source and target domains, including: For each spatial pixel and each time sub-period, the effective sample size of each observation channel in the corresponding source domain training sample set and target domain training sample set is statistically analyzed to determine the minimum effective sample size, and the corresponding quantile set is selected from multiple candidate quantile sets based on the minimum effective sample size. Based on the selected set of quantiles, the quantile values ​​of the source domain training sample set and the target domain training sample set are calculated in each observation channel. Through interpolation between adjacent quantile values ​​and tail fitting outside the boundary quantile values, the empirical cumulative distribution function of the source domain and the empirical cumulative distribution function of the target domain are established respectively. Using the source domain empirical cumulative distribution function and the target domain empirical cumulative distribution function, the source domain training samples and target domain training samples of each observation channel are transformed into probability spaces to obtain source domain probability samples and target domain probability samples.

5. The method for consistency correction of multi-channel satellite vegetation optical thickness data as described in claim 4, characterized in that, Interpolation between adjacent quantiles and tail fitting outside of boundary quantiles include: For training samples located between the second lowest quantile and the second highest quantile, linear interpolation is used to determine the cumulative probability. For training samples that are below the second lowest quantile or above the second highest quantile, a linear function is fitted using the ranking value of the corresponding tail training samples and the tail cumulative probability. When the number of training samples at the tail is less than the tail threshold, the boundary probability corresponding to the tail threshold is used instead of the cumulative tail probability for a linear function fitting.

6. The method for consistency correction of multi-channel satellite vegetation optical thickness data as described in claim 4, characterized in that, Mapping the cumulative probability data to Gaussian space, estimating and positive-definiteizing the correlation matrices of the source and target domains, and constructing the connectivity structure transition matrix using matrix decomposition include: From all the source domain probability samples and target domain probability samples, select the source domain joint effective samples and target domain joint effective samples where the vegetation optical thickness values ​​obtained by each observation channel under the same spatial pixel are all effective values. By using the inverse function of the standard normal cumulative distribution function, the joint effective samples of the source domain and the joint effective samples of the target domain are transformed into Gaussian space to obtain Gaussian samples of the source domain and Gaussian samples of the target domain. Based on the source domain Gaussian samples and the target domain Gaussian samples, the source domain channel correlation matrix and the target domain channel correlation matrix are determined respectively, and the two channel correlation matrices are positive definite. Matrix decomposition is performed on the positive definite source domain channel correlation matrix and the target domain channel correlation matrix respectively to obtain the source domain decomposition matrix and the target domain decomposition matrix; Based on the source domain decomposition matrix and the target domain decomposition matrix, a connection structure migration matrix is ​​constructed to transform the source domain channel-related structure into the target domain channel-related structure.

7. The method for consistency correction of multi-channel satellite vegetation optical thickness data as described in claim 1, characterized in that, Based on the source domain decomposition matrix and the target domain decomposition matrix, a connection structure transfer matrix is ​​constructed to transfer the source domain channel-related structure to the target domain channel-related structure. The corresponding formula is: A=C 目标 C 源 -1 , Where A is the connection structure transition matrix; C 源 -1 It is the inverse matrix of the source domain decomposition matrix; C 目标 It is the target domain decomposition matrix.

8. The method for consistency correction of multi-channel satellite vegetation optical thickness data as described in claim 1, characterized in that, The method further includes: Calculate the daily average value of the source domain multi-channel vegetation optical thickness time series data before correction and the daily average value of the source domain multi-channel vegetation optical thickness time series data after correction, respectively; The daily average values ​​of the source domain multi-channel vegetation optical thickness time series data before and after correction were compared with the daily average values ​​of the target domain multi-channel vegetation optical thickness time series data to verify the corrected source domain multi-channel vegetation optical thickness time series data.

9. A multi-channel satellite vegetation optical thickness data consistency correction system, characterized in that, include: The data acquisition and preprocessing module is used to acquire and preprocess multi-channel vegetation optical thickness time-series data from source domain sensors and target domain sensors. The adaptive empirical distribution construction module is used to calculate the effective sample size of each observation channel based on the preprocessed multi-channel vegetation optical thickness time series data to adaptively select quantiles, and construct the continuous cumulative distribution functions of the source domain and the target domain respectively through quantile interpolation and tail extrapolation, thereby obtaining the cumulative probability data of the source domain and the target domain. The dependent structure transfer module is used to map the cumulative probability data to Gaussian space, estimate and positive definite the correlation matrix between the source domain and the target domain, and construct the connection structure transfer matrix using matrix decomposition; The inverse cumulative distribution mapping module is used to perform source domain cumulative distribution mapping, joint dependency structure correction and target domain inverse cumulative distribution mapping sequentially on the source domain multi-channel vegetation optical thickness time series data to be corrected, based on the continuous cumulative distribution function of the source domain and the connection structure migration matrix of the target domain, to obtain the corrected source domain multi-channel vegetation optical thickness time series data.

10. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium stores computer instructions that cause the computer to execute the multi-channel satellite vegetation optical thickness data consistency correction method according to any one of claims 1 to 8.

Images