Adaptive Channel Selection Method and System for Hyperspectral Infrared Satellite Data Assimilation

Abstract

The present invention discloses an adaptive channel selection method and system for hyperspectral infrared satellite data assimilation, including an input module, an observation impact on background error calculation module, an assimilated observation rejection module, and a posterior ensemble acquisition module. Given a set of atmospheric forecast ensembles and observations; before assimilation, in the observation space, calculate the ratio of the posterior background error covariance to the prior background error covariance (Pa / Pb), and assimilate the observation data in ascending order of this value. For the observation with the smallest Pa / Pb in each iteration, it will be assimilated; after assimilating a certain observation, use the posterior ensemble as the new prior ensemble, and reject the assimilated observations from the observations; when the Pa / Pb of the remaining observations is greater than a preset threshold, the assimilation terminates; after assimilating all the observations within the Pa / Pb threshold, obtain the posterior ensemble and calculate the ensemble mean. The present invention is not only suitable for the assimilation of rapidly changing weather, but also applicable to hyperspectral satellite data with a wider spectral coverage width.

Description

Technical Field

[0001] The present invention relates to an adaptive channel selection method and system for hyperspectral infrared satellite data assimilation, belonging to the technical field of atmospheric analysis. Background Art

[0002] The high vertical resolution atmospheric information provided by satellite infrared hyperspectral sounders makes an important contribution to the accuracy of numerical weather prediction (NWP). However, there are numerous channels in hyperspectral sounders, with a large amount of data redundancy and correlation between channels. In practical applications, data dimensionality reduction and removal of observational correlation must be carried out. In addition, due to limitations in computing power and assimilation timeliness, it is difficult to assimilate all channels in NWP applications. Regarding data dimensionality reduction and channel selection for hyperspectral data, there have been many studies. Among them, the entropy reduction (ER) channel selection method based on the Shannon (1998) measure of source uncertainty and the principal component analysis (PCA) method based on spatial transformation are the most widely used.

[0003] However, the ER-based channel selection discards a large number of channels, significantly reducing the original spectral coverage width. In addition, the ER channel selection method is based on the linear theory of optimal estimation and usually fails to consider flow-dependent and time-varying optimal channel selection. With the improvement of numerical models and observational resolutions, more details of multi-scale weather features and related evolution processes can be revealed. Therefore, the channels selected by the ER method may not be the optimal solution. Similar to the ER-based method, the PCA-based method also fails to consider the time variation of the selected principal components. And as a linear transformation, the leading principal components of the PCA channel selection method may lose the non-linear information of the original data. Moreover, the PCA-based channel selection method requires the development of PCA-Based radiative transfer models, such as PC_RTTOV (Matricardi 2010) or PCRTM (Liu et al. 2006). Therefore, how to select the optimal channel data at the assimilation moment for effective hyperspectral data assimilation or inversion has always been a key issue in hyperspectral data assimilation. Summary of the Invention

[0004] Object of the Invention: Aiming at the deficiencies of the ER channel selection technology, the present invention provides an adaptive channel selection method and system suitable for assimilation of rapidly changing weather and also applicable to hyperspectral satellite data with a wider spectral coverage width.

[0005] Technical Solution: To achieve the above object, the technical solution adopted by the present invention is as follows:

[0006] An adaptive channel selection method for hyperspectral infrared satellite data assimilation proposed by the present invention first gives a set of atmospheric forecast ensembles, and the ensemble members are P b is the prior background error covariance of the ensemble, and P a is the posterior background error covariance of the analysis field after assimilating the observations. In the assimilation cycle, the ratio (Pa / Pb) of the posterior background error covariance to the prior background error covariance is used to estimate to what extent each observation (if assimilated individually) will reduce the ensemble variance of the system state variables. If the expected posterior error covariance of the assimilated observation is very close to the prior background error covariance, that is, the ratio of Pa / Pb is close to 1, it can be expected that the impact of this observation on the system is very small, so this observation can be skipped. Setting a smaller Pa / Pb threshold can ensure that only the observations that have a large impact on the system variance are assimilated, so as to achieve adaptive channel selection. Thus, the adaptive channel selection method for hyperspectral infrared satellite data assimilation provided by the present invention provides a flow-dependent and time-varying channel selection algorithm in the assimilation process by capturing the reduction of the background error that changes with time, including the following steps:

[0007] Step 1, give a set of atmospheric forecast ensembles and observations;

[0008] Step 2, calculate the impact of the observations on the prior background error covariance before assimilation:

[0009] Step 2.1, in the observation space, calculate the ratio η of the posterior background error covariance obtained from each observation to the prior background error covariance;

[0010] Step 2.2, assimilate the observation data in the increasing order of the ratio η calculated in Step 2.1, and the observation with the smallest ratio η in each iteration will be assimilated;

[0011] Step 3, set the threshold η' of the ratio η, and jump out of the assimilation cycle when it is greater than the threshold η';

[0012] After assimilating a certain observation, use the posterior ensemble as the new prior ensemble. After removing the assimilated observations from the observations, repeat Step 2; when the ratio η of the posterior background error covariance to the prior background error covariance of the remaining observations is greater than the preset threshold η', it indicates that the unassimilated observations only reduce the ensemble dispersion and cannot significantly improve the analysis result, and the assimilation terminates;

[0013] Step 4, obtain the posterior ensemble after assimilation:

[0014] After assimilating all the observations within the threshold η', obtain the posterior ensemble and calculate the ensemble mean.

[0015] Preferably, the ratio η of the posterior background error covariance to the prior background error covariance:

[0016]

[0017] where Pa / Pb is the ratio of the posterior background error covariance to the prior background error covariance, P a is the posterior background error covariance in the observation space, P b is the prior background error covariance in the observation space, h is the forward operator for a certain observation; for a certain observation, is its observation error, is the variance of its prior background error.

[0018] Preferably, in step 1, a set of atmospheric forecast ensembles i = 1,..., N, where N is the size of the atmospheric forecast ensemble, is denoted as the ensemble mean, and the ensemble perturbation of the i-th ensemble member is Write the ensemble perturbation in the form of the square root of the background error covariance Then the background error covariance is expressed as P b = XX T .

[0019] Preferably, in step 1, an observation y with an error covariance matrix R is given. Assuming that the observation errors are uncorrelated, R is a diagonal matrix, and its diagonal elements are the error variances of the observed variables.

[0020] Preferably, assimilate and update the prior ensemble mean, and the updated posterior ensemble mean is given by the following formula:

[0021]

[0022] where ρ is the localization matrix, ο represents the Schur product, H represents the forward observation operator, represents the prior mean estimate in the observation space.

[0023] Preferably, assimilate and update the ensemble perturbation. The i-th ensemble perturbation is updated through the following equation:

[0024]

[0025] where, represents the posterior perturbation of the i-th ensemble member.

[0026] Another technical object of the present invention is to provide an adaptive channel selection system for hyperspectral infrared satellite data assimilation, including an input module, an observation impact on background error calculation module, an eliminated assimilated observation module, and a posterior ensemble obtaining module, wherein:

[0027] The input module is used to input a set of atmospheric forecast ensembles and observations;

[0028] The module for calculating the influence of observations on background error is used to calculate the ratio η of the posterior background error covariance to the prior background error covariance in the observation space; the observation data is assimilated in ascending order of the ratio η, and for each iteration, the observation with the smallest ratio η will be assimilated;

[0029] The module for removing assimilated observations is used to use the posterior ensemble as the new prior ensemble after assimilating a certain observation, remove the assimilated observations from the observations, and then return to the module for calculating the influence of observations on background error; when the ratio η of the posterior background error covariance to the prior background error covariance of the remaining observations is greater than the preset threshold η', the assimilation terminates;

[0030] The module for obtaining the posterior ensemble is used to obtain the posterior ensemble and calculate the ensemble mean after assimilating all observations within the threshold η'.

[0031] Preferably: the formula for calculating the ratio of the posterior background error covariance to the prior background error covariance in the module for calculating the influence of observations on background error is:

[0032]

[0033] where Pa / Pb is the ratio of the posterior background error covariance to the prior background error covariance, P a is the posterior background error covariance in the observation space, P b is the prior background error covariance in the observation space, h is the forward operator of a certain observation; for a certain observation, is its observation error, is the variance of its prior background error.

[0034] Compared with the prior art, the present invention has the following beneficial effects:

[0035] Since the background field changes with time, the observations selected by Pa / Pb have a flow-dependent characteristic, so they are more suitable for the assimilation of rapidly changing weather; at the same time, since there is no direct rejection of channels, the spectral coverage width of the hyperspectral detector is retained to the greatest extent. Description of the Drawings

[0036] Figure 1 It is a schematic flowchart of the adaptive channel selection method (abbreviated as Pa / Pb method) for hyperspectral infrared satellite data assimilation described in the present invention.

[0037] Figure 2Selected channels and associated radiation observations for different assimilation times (black for time 040030, gray for time 040100), based on the ER method (cumulative ER ratio equal to 0.95) and the Pa / Pb (threshold of 0.98) method.

[0038] Figure 3 Comparison diagrams of the average RMS errors of different state variables in the analysis field after assimilation using the ER method and the Pa / Pb method under ClearSky; in the diagrams: (a) represents the comparison diagram of the average RMS errors of temperature in the 7.5 km horizontal resolution assimilation experiment in the analysis field after assimilation using the ER method and the Pa / Pb method; (b) represents the comparison diagram of the average RMS errors of temperature in the 1.5 km horizontal resolution assimilation experiment in the analysis field after assimilation using the ER method and the Pa / Pb method; (c) represents the comparison diagram of the average RMS errors of temperature in the 300 m horizontal resolution assimilation experiment in the analysis field after assimilation using the ER method and the Pa / Pb method; (d) represents the comparison diagram of the average RMS errors of specific humidity in the 7.5 km horizontal resolution assimilation experiment in the analysis field after assimilation using the ER method and the Pa / Pb method; (e) represents the comparison diagram of the average RMS errors of specific humidity in the 1.5 km horizontal resolution assimilation experiment in the analysis field after assimilation using the ER method and the Pa / Pb method; (f) represents the comparison diagram of the average RMS errors of specific humidity in the 300 m horizontal resolution assimilation experiment in the analysis field after assimilation using the ER method and the Pa / Pb method; (g) represents the comparison diagram of the average RMS errors of the U component of the wind field in the 7.5 km horizontal resolution assimilation experiment in the analysis field after assimilation using the ER method and the Pa / Pb method; (h) represents the comparison diagram of the average RMS errors of the U component of the wind field in the 1.5 km horizontal resolution assimilation experiment in the analysis field after assimilation using the ER method and the Pa / Pb method; (i) represents the comparison diagram of the average RMS errors of the U component of the wind field in the 300 m horizontal resolution assimilation experiment in the analysis field after assimilation using the ER method and the Pa / Pb method; (j) represents the comparison diagram of the average RMS errors of the V component of the wind field in the 7.5 km horizontal resolution assimilation experiment in the analysis field after assimilation using the ER method and the Pa / Pb method; (k) represents the comparison diagram of the average RMS errors of the V component of the wind field in the 1.5 km horizontal resolution assimilation experiment in the analysis field after assimilation using the ER method and the Pa / Pb method; (l) represents the comparison diagram of the average RMS errors of the V component of the wind field in the 300 m horizontal resolution assimilation experiment in the analysis field after assimilation using the ER method and the Pa / Pb method; in each diagram, circles represent cumulative ER values reaching 0.80, 0.85, 0.90, and 0.95 respectively; crosses represent Pa / Pb thresholds of 0.90, 0.95, 0.96, 0.97, 0.98, and 0.99; plus signs represent assimilating all channels. Detailed implementation

[0039] The present invention will be further illustrated below in conjunction with the accompanying drawings and specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and not to limit the scope of the present invention. After reading the present invention, those skilled in the art's various equivalent modifications of the present invention all fall within the scope defined by the appended claims of this application.

[0040] An adaptive channel selection method for hyperspectral infrared satellite data assimilation proposed by the present invention first needs to calculate the ensemble-based background error covariance in the framework of the ensemble Kalman filter, and then estimate to what extent each observation will reduce the system variance through the ratio of the posterior background error covariance to the prior background field error covariance (Pa / Pb). Subsequently, the observations are sorted in ascending order of Pa / Pb (in the order of decreasing degree of background error reduction), and a Pa / Pb threshold is set so that only the observations that have a greater impact on the background error are assimilated. Finally, assimilation experiments are carried out using the channels selected by the adaptive method.

[0041] Specifically, as Figure 1 shown, an adaptive channel selection method for hyperspectral infrared satellite data assimilation proposed by the present invention includes the following steps:

[0042] Step 1, given a set of atmospheric forecast ensembles and observations;

[0043] Step 1.1, given a set of atmospheric forecast ensembles;

[0044] Given a set of atmospheric forecast ensembles with a size of N i = 1,..., N, denoted as the ensemble mean, the i-th ensemble member has an ensemble perturbation of For the convenience of calculating the background error covariance during assimilation, the ensemble perturbation can be written in the form of the square root of the background error covariance Then the background error covariance is denoted as P b = XX T ;

[0045] Step 1.2, given the observations to be assimilated;

[0046] Given observations y with an error covariance matrix R. Generally, it is assumed that the observation errors are uncorrelated, and R is a diagonal matrix, and its diagonal elements are the error variances of the observed variables; in this ideal experiment, it is assumed that the error standard deviations of different channels of the satellite are the same and are fixed values (1.5K). Use the ensemble square root filter (EnSRF; Whitaker and Hamill 2002) to assimilate the observations y, and update the posterior ensemble mean and perturbation respectively; the updated posterior ensemble mean is given by the following formula:

[0047]

[0048] Among them, ρ is the localization matrix, ο represents the Schur product, and H represents the forward observation operator. represents the prior mean estimate of the observation space;

[0049] The i-th ensemble perturbation is updated through the following equation:

[0050]

[0051] Among them, represents the i-th posterior ensemble perturbation;

[0052] Step 2: Calculate the impact of the observation on the background error before assimilation;

[0053] Step 2.1: Calculate the magnitude of the reduction in background error;

[0054] As a sequential assimilation filter, EnSRF can assimilate observation data one by one. When the observation errors are uncorrelated, the sequential EnSRF is equivalent to the matrix version. For a single observation to be assimilated, the observation error covariance matrix R is a scalar, denoted as And the estimated background error covariance matrix HP of the observation space b H T is also a scalar, denoted as h is the forward operator of a certain observation (i.e., a row vector of the forward observation operator H). Therefore, in the observation space, the ratio of the posterior background error covariance to the prior background error covariance can be written as:

[0055]

[0056] Step 2.2: Sort and assimilate the observations;

[0057] Since the observations are not assimilated in the order of input of the observation data, but in the increasing order of the ratio Pa / Pb of the posterior background error covariance to the prior background error covariance. Therefore, for the observation with the smallest Pa / Pb in each iteration, it will be assimilated;

[0058] Step 3: Set the threshold η' of the ratio η, and jump out of the assimilation loop when it is greater than the threshold η';

[0059] After assimilating a certain observation, use the posterior ensemble as the new prior ensemble. After removing the assimilated observation from the observations, repeat Step 2; when the ratio η of the posterior background error covariance to the prior background error covariance of the remaining observations is greater than the preset threshold η', it indicates that the unassimilated observations only reduce the ensemble dispersion and cannot significantly improve the analysis result, and the assimilation terminates;

[0060] Step 4: Obtain the assimilated posterior ensemble.

[0061] After assimilating all the observations within the threshold η’, obtain the posterior ensemble and calculate the ensemble mean.

[0062] Based on the above-mentioned adaptive channel selection method for hyperspectral infrared satellite data assimilation, the present invention proposes an adaptive channel selection system for hyperspectral infrared satellite data assimilation, including an input module, an observation impact on background error calculation module, a removed assimilated observation module, and a posterior ensemble obtaining module, where:

[0063] The input module is used to input a set of atmospheric forecast ensembles and observations.

[0064] The observation impact on background error calculation module is used to calculate the ratio (Pa / Pb) of the posterior background error covariance to the prior background error covariance in the observation space. The observation data is assimilated in ascending order of Pa / Pb, and the observation with the smallest Pa / Pb in each iteration will be assimilated.

[0065] The removed assimilated observation module is used to use the posterior ensemble as the new prior ensemble after assimilating a certain observation, remove the assimilated observation from the observations, and then return to the observation impact on background error calculation module. When the Pa / Pb of the remaining observations is greater than the preset threshold, the assimilation terminates.

[0066] The posterior ensemble obtaining module is used to obtain the posterior ensemble and calculate the ensemble mean after assimilating all the observations within the Pa / Pb threshold.

[0067] In this embodiment, the ER channel selection method and the adaptive channel selection method (Pa / Pb) are respectively used to assimilate the multi-channel satellite infrared radiation data (FY-4AGIIRS), and the root mean square (RMS) error of the analysis field mean after assimilation is examined under different weather conditions (clear sky or all sky) to evaluate the assimilation effect. The horizontal localization scale sizes adopted in the experiments with different model resolutions are different and have been adjusted to the optimal. The specific experimental settings are shown in Table 1.

[0068] Table 1 Parameter configuration of assimilation experiments with different model resolutions

[0069]

[0070] Figure 2Shows the variation of the observations selected by the Pa / Pb method with a cumulative ER of 95% and a threshold of 0.98 over time under AllSky (Full Sky). Although the quantity of observations selected by the ER method has changed slightly over time due to quality control reasons, the channels it selects are fixed. In contrast, the adaptive method can capture the impact of radiation data on the background error covariance at different assimilation times, thus providing flow-dependent and time-varying channel selection. And compared with the ER method, the vertical distribution range of the observations selected by the Pa / Pb method is wider.

[0071] Figure 3 Is the comparison of the average RMS of different state variables in the analysis fields after assimilation using the ER and Pa / Pb methods under ClearSky (Clear Sky) conditions. Under clear sky, when the quantity of assimilated observations is comparable, the average RMS of Pa / Pb in variables such as temperature, specific humidity, U, and V is generally smaller than that of ER. The performance under AllSky (figure omitted) is similar to that under clear sky. And when the model resolution is increased from 7.5 km to 300 m, the results are consistent, that is, the adaptive channel selection method has better analysis field results.

[0072] Compared with the traditional channel selection method based on information entropy, the adaptive channel selection method proposed in this embodiment can capture the reduction of the time-varying background error, thus providing flow-dependent and time-varying channel selection. Compared with the analysis field generated after selecting the channels by assimilating information entropy, the assimilation using the adaptive channel selection method has a smaller RMS error of state variables and provides a better analysis field for tropical cyclone forecasting at horizontal resolutions from kilometers to sub-kilometers.

[0073] The above are only the preferred embodiments of the present invention. It should be noted that: for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements can still be made, and these improvements and refinements should also be regarded as the protection scope of the present invention.

Claims

1. An adaptive channel selection method for hyperspectral infrared satellite data assimilation, characterized in that ,By capturing the reduction of the time-varying background error, a flow-dependent ,time-varying channel selection algorithm is provided in the assimilation process, which includes the following steps: Step 1: Given a set of atmospheric forecasts and observations; Step 2, calculate the effect of observations on the prior background error covariance before assimilation: Step 2.1, in the observation space, calculate the ratio η of the posterior background error covariance to the prior background error covariance for each observation; Step 2.2, assimilate the observation data in the increasing order of the ratio η calculated in step 2.1, and the observation with the smallest ratio η in each iteration will be assimilated; Step 3, set the threshold η' of the ratio η, and exit the assimilation cycle when it is greater than the threshold η': After assimilating an observation, use the posterior set as the new prior set, remove the assimilated observations from the observations, and repeat step 2; when the ratio η of the posterior background error covariance of the remaining observations to the prior background error covariance is greater than the preset threshold η', it indicates that the unassimilated observations only reduce the set dispersion but cannot significantly improve the analysis results, and the assimilation is terminated; Step 4, obtain the assimilated posterior set: After assimilating all observations within the threshold η', the posterior set is obtained and the ensemble mean is calculated.

2. The adaptive channel selection method for hyperspectral infrared satellite data assimilation according to claim 1, characterized in that: The ratio of the posterior background error covariance to the prior background error covariance η: Where Pa / Pb is the ratio of the posterior background error covariance to the prior background error covariance, P a is the posterior background error covariance in the observation space, P b is the prior background error covariance of the observation space, h is the forward operator of a certain observation; for a certain observation, is the observation error, is the variance of the prior background error.

3. The adaptive channel selection method for hyperspectral infrared satellite data assimilation according to claim 2 is characterized in that: In step 1, a set of atmospheric forecasts is given N is the atmospheric forecast ensemble size, Expressed as the ensemble average, the i-th ensemble member The collective perturbation of The ensemble perturbation is written as the square root of the background error covariance Then the prior background error covariance is expressed as P b =XX T .

4. The adaptive channel selection method for hyperspectral infrared satellite data assimilation according to claim 3 is characterized by: In step 1, given the observation y with error covariance matrix R, assuming that the observation errors are uncorrelated, R is a diagonal matrix whose diagonal elements are the error variances of the observed variables.

5. The adaptive channel selection method for hyperspectral infrared satellite data assimilation according to claim 4 is characterized by: The average of the prior set is updated by assimilation, and the average of the updated posterior set is Given by: where ρ is the localization matrix, represents the Schur product, H represents the forward observation operator, represents the a priori mean estimate of the observation space.

6. The adaptive channel selection method for hyperspectral infrared satellite data assimilation according to claim 5, characterized in that: The collective perturbation is assimilated and updated, and the i-th collective perturbation is updated by the following equation: in, represents the posterior perturbation of the i-th set member.

7. An adaptive channel selection system for hyperspectral infrared satellite data assimilation, characterized by: It includes input module, calculation module of observation impact on background error, module for removing assimilated observations, and module for obtaining posterior set, among which: The input module is used to input a set of atmospheric forecast sets and observations; The module for calculating the influence of observations on background errors is used to calculate the ratio η of the posterior background error covariance to the prior background error covariance in the observation space; assimilate the observation data in the increasing order of the ratio η, and the observation with the smallest ratio η in each iteration will be assimilated; The module for removing assimilated observations is used to assimilate a certain observation, use the posterior set as a new prior set, remove the assimilated observations from the observations, and return to the module for calculating the impact of the observations on the background error; when the ratio η of the posterior background error covariance of the remaining observations to the prior background error covariance is greater than a preset threshold η', the assimilation is terminated; The posterior set acquisition module is used to obtain the posterior set and calculate the ensemble average after assimilating all observations within the threshold η'.

8. The adaptive channel selection system for hyperspectral infrared satellite data assimilation according to claim 7, characterized in that: The calculation formula for the ratio of the posterior background error covariance to the prior background error covariance in the observation effect calculation module on the background error is: Where Pa / Pb is the ratio of the posterior background error covariance to the prior background error covariance, P a is the posterior background error covariance in the observation space, P b is the prior background error covariance of the observation space, h is the forward operator of a certain observation; for a certain observation, is the observation error, is the variance of the prior background error.

Images