Chlorophyll concentration remote sensing data product interpolation method

Through the adaptive spatiotemporal tensor completion method, the tensor truncated kernel norm model is used in conjunction with time, space and periodic characteristics, the uncertainty problem of interpolation of chlorophyll a concentration in the existing technology is solved, and the problem of uncertainty in interpolation of remote sensing data in chlorophyll a concentration is realized. High-precision and low-cost data filling improves the application capabilities of marine remote sensing data.

CN119089121APending Publication Date: 2024-12-06NANJING GUOJI SCI & TECH RES INST CO LTD
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202411068607.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-06
Publication Date
2024-12-06

AI Technical Summary

Technical Problem

When using spatiotemporal information to interpolate chlorophyll a concentration remote sensing data, the prior art restricts the insertion of the interpolation of severe data loss and severe data loss and the high-variation areas. Applicability of supplement.

Method used

Adaptive spatiotemporal tensor completion method is adopted to rearrange the remote sensing data of chlorophyll a concentration into three-dimensional spatiotemporal tensors, and adaptive spatiotemporal tensor completion model is constructed based on the tensor truncated kernel norm model, using time, space and periodicity Features are interpolated to realize multi-dimensional feature extraction of data and accurate filling of missing data.

Benefits of technology

It improves the accuracy and applicability of interpolated data, can effectively fill the remote sensing time series of chlorophyll a concentration in severe absence, and does not require additional data training. It is suitable for remote sensing data interpolation tasks in different spatial ranges, reducing application costs.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119089121A_ABST
    Figure CN119089121A_ABST
Patent Text Reader

Abstract

The invention relates to the technical field of remote sensing data processing, and discloses a chlorophyll a concentration remote sensing data product interpolation method, which comprises the following steps of: rearranging a three-dimensional tensor x formed by an original chlorophyll a concentration remote sensing time sequence to be interpolated into a new three-dimensional space-time tensor form x'according to space, time and periodic characteristics; constructing an adaptive space-time tensor completion model, inputting a space-time tensor x ', and solving the tensor completion model to obtain a solution of the space-time tensor x'; and carrying out inverse transformation on the solution of the space-time tensor x'and then converting the solution into a format corresponding to original data to obtain an interpolated chlorophyll a concentration remote sensing time sequence. According to the method, a self-adaptive space-time tensor complementation method is constructed for filling the missing region, a tensor form with a lower rank characteristic can be obtained, data does not need to be additionally collected for training while the interpolation precision is ensured, and the practical application capability of a seriously missing satellite data product in marine environment monitoring is greatly improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to the technical field of remote sensing data processing, and in particular to a method for interpolating chlorophyll a concentration remote sensing data products. Background Art

[0002] Chlorophyll a concentration (Chla) is an important indicator of the amount of phytoplankton in a water body. Therefore, Chla data are widely used in marine ecosystem production assessment, phytoplankton phenology detection, and algal bloom monitoring. In recent years, satellite remote sensing has become a widely used method for monitoring global Chla. Satellite sensors can cover large areas with temporal resolutions ranging from daily to monthly.

[0003] However, due to cloud cover, sunlight pollution and other unfavorable conditions, the Chla data products obtained by satellites are inevitably missing, which greatly hinders their further application in phytoplankton-related fields. For example, more than 70% of the data in the global data products of MODIS-Terra / Aqua and VIIRS-SNPP satellites are missing. Therefore, reconstructing the missing information and obtaining the interpolated Chla time series products can better characterize the actual changing state of marine phytoplankton over a large range and over a long period of time.

[0004] Currently, many interpolation methods have been proposed to reconstruct missing information in Chla remote sensing time series products. Current methods can be generally divided into data-driven and non-data-driven. Among them, data-driven methods mainly include traditional machine learning and deep learning methods, which fill in the missing data of Chla by establishing an empirical relationship between missing data and valid or auxiliary data. Typical traditional machine learning methods include self-organizing maps, support vector regression, etc. Deep learning methods, such as data interpolation convolutional autoencoders (DINCAE), convolutional neural networks, and spatiotemporal attention generative adversarial networks, have also been introduced into the Chla interpolation task due to their data feature extraction and nonlinear modeling capabilities. Overall, data-driven methods have achieved good results in filling in missing data in Chla, but their generalization ability is always limited by the scale and spatiotemporal diversity of the training dataset. When there are serious missing data in large-scale remote sensing time series data, it is very difficult for data-driven methods to collect large-scale datasets, and the applicability of the methods is limited.

[0005] In contrast, non-data-driven interpolation methods are mainly based on the idea of ​​interpolation, and only need to use the spatial, temporal or spatiotemporal characteristics of the original Chla data itself to restore missing data. This method has strong generalization and application capabilities and has been widely used in the field of marine remote sensing. Classic methods include linear interpolation, optimal interpolation (OI) and Kriging algorithm. These methods are simple and efficient, but can only use the temporal information or spatial information of a single pixel. When there are serious temporal or spatial missing, this limitation will cause abnormal spatiotemporal patterns in the interpolation results. Therefore, methods that use spatial domain information to assist temporal interpolation have been proposed. For example, the data interpolation empirical orthogonal function (DINEOF) method uses spatiotemporal data features to achieve more accurate reconstruction of missing data. This method has been effectively applied to interpolation tasks of various types of marine remote sensing data. In addition, some scholars have used spatial neighboring pixels with similar temporal changes to fill the missing time series of target pixels. However, the current methods of interpolation using spatiotemporal information still have large uncertainties in areas with severe data missing and high variation. This is mainly because non-data-driven methods still do not fully utilize temporal and spatial information when filling the gaps in Chla time series.

[0006] Existing methods tend to pre-fill missing data in the temporal domain by temporal linear interpolation, which can only use the confidence of temporally adjacent observations and ignore the unique periodic characteristics of phytoplankton growth, that is, Chla should be similar in the same period in different years. In the spatial domain, although there are a few spatiotemporal interpolation methods that reconstruct missing pixels by using spatially adjacent pixels, they usually apply spatial and temporal information separately in different steps and cannot be used synergistically. In fact, the long-term Chla time series accumulated over many years can provide rich reference information for filling in the missing data, including spatial, temporal and periodic characteristics. However, current methods still fail to effectively coordinate this information to reconstruct long-term Chla time series products. This limits the applicability of existing methods in areas with severe data gaps, especially the average missing rate of daily Chla products worldwide exceeds 70%, and is as high as 80% locally. Therefore, developing a spatiotemporal interpolation method that can simultaneously extract temporal, spatial and periodic variation characteristics can provide an effective solution for interpolating severe missing data provided by Chla remote sensing time series products. Summary of the invention

[0007] The purpose of the present invention is to provide a chlorophyll a concentration remote sensing data product interpolation method based on adaptive spatiotemporal tensor completion, which utilizes the temporal, spatial and periodic characteristics of the chlorophyll a concentration remote sensing data product itself to achieve interpolation of large-scale missing data and improve the application capability of the data product in marine ecological environment monitoring.

[0008] In order to achieve the above object, the present invention adopts the following technical solutions:

[0009] A method for interpolating chlorophyll a concentration remote sensing data products, comprising the following steps:

[0010] Step A: A three-dimensional tensor formed from the original chlorophyll a concentration remote sensing time series to be interpolated Rearranged into a new three-dimensional space-time tensor form according to the characteristics of space, time and period

[0011] Step B: Construct an adaptive spatiotemporal tensor completion model based on the tensor truncated nuclear norm model, and input the rearranged chlorophyll a concentration spatiotemporal tensor The tensor completion model is solved by adaptively allocating tensor weights in different dimensions to obtain the spatiotemporal tensor. The solution;

[0012] Step C: Transform the space-time tensor The solution is transformed into the format corresponding to the original data after the rearrangement in step A, and the interpolated chlorophyll a concentration remote sensing time series is obtained.

[0013] Preferably, the specific steps in step A are as follows:

[0014] A1. Arrange the original chlorophyll a concentration remote sensing time series images in chronological order into a three-dimensional tensor, recorded as:

[0015] Where m and n represent the size of the image, and t represents the length of the date in the entire time series;

[0016] A2. Since the concentration of chlorophyll a has the characteristic of periodic change, the original tensor is rearranged into a new three-dimensional tensor, recorded as:

[0017] Where ny represents the length of the entire time series in years, and nd is the number of observations per year.

[0018] Preferably, the specific implementation process of constructing the adaptive spatiotemporal tensor completion model is as follows:

[0019] B11. Functional expression of the rank-minimization tensor completion problem:

[0020]

[0021] Where rank is the algebraic rank of the chlorophyll a concentration remote sensing time series data tensor, Ω represents the index set of the observed pixels, and P Ω represents the orthogonal projection on Ω;

[0022] B12. Using the weighted convex kernel norm to approximate the tensor rank, the above formula is converted to:

[0023]

[0024] In the formula, Representing a tensor The kth expansion matrix of k yes The non-negative weights of ||·|| * represents the nuclear norm regularization term;

[0025] B13. Based on low-rank matrix completion and using tensor trace norm instead of matrix nuclear norm, the low-rank tensor completion problem in the previous step is converted to:

[0026]

[0027] Where Y represents the missing tensor, A is a diagonal matrix with diagonal elements of 0 and 1;

[0028] When a pixel in Y is missing, the value is 0, otherwise it is 1;

[0029] is the fidelity term of the data, is a regularization term, which represents the prior distribution of the real data, and λ is a regularization parameter, which is used to balance the contribution of the above two items.

[0030] Preferably, the specific process of solving the tensor completion model by adaptively allocating tensor weights in different dimensions is as follows:

[0031] B21. Rearranged 3D Tensors The k-th dimension expansion matrix is ​​obtained by singular value decomposition method Singular values ​​of and arrange in descending order Construct a one-dimensional singular value sequence;

[0032] B22. Connect the first point in the sequence to the maximum value to form a straight line, subtract each point in the sequence from the straight line, and then normalize the maximum and minimum values ​​to obtain a new singular value difference sequence

[0033] B23. Calculate the low-rank index for each dimension where i k-max is the local maximum of the singular value difference sequence, q k is the length of the singular value sequence in k dimensions;

[0034] B24. Calculate the weight index of different dimensions k of the tensor in

[0035] Preferably, the obtaining of the space-time tensor The specific process of the solution is:

[0036] B31. After obtaining the weights of different expansion dimensions k, introduce a new tensor value variable M k Replace X in step B13 k , and obtain the new model solution formula:

[0037]

[0038] Further define a Lagrangian function:

[0039]

[0040] In the formula, B k is the Lagrangian variable, β is the penalty parameter;

[0041] B33. Based on the above steps, define the optimization sub-problem to be solved by the model:

[0042]

[0043] B34, solve the sub-problem M k The formula for X is:

[0044]

[0045] B35. Set the number of iterations i and update the Lagrangian variable B k In addition, the weight parameter α in each iteration k are updated by solving the tensor completion model. k The update formula is as follows:

[0046]

[0047] Preferably, the solution obtained in step B According to step A, the rearrangement process is reversed to The interpolated chlorophyll a concentration remote sensing time series is obtained.

[0048] Compared with the prior art, the present invention has the following beneficial effects:

[0049] The present invention introduces the tensor completion concept into the interpolation of missing data of chlorophyll a concentration remote sensing time series. By integrating multidimensional spatiotemporal information, including spatial range, interannual variation and periodic characteristics, a low-rank three-dimensional spatiotemporal tensor more suitable for multidimensional feature extraction is constructed. Then, an adaptive spatiotemporal tensor completion model is constructed based on the tensor truncated nuclear norm model to achieve the interpolation of missing data of chlorophyll a concentration remote sensing time series. Compared with previous methods, this method can be applied to remote sensing data interpolation tasks of different spatial ranges without data training, with low application cost, high accuracy and strong applicability. It can not only cope with the interpolation of chlorophyll a concentration remote sensing time series under severe missing data, but also can be extended to the interpolation tasks of other marine remote sensing data products. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 This is a flow chart of the interpolation method for chlorophyll a concentration remote sensing data products of the present invention;

[0051] Figure 2 It is a schematic diagram of the tensor rearrangement of the remote sensing time series of chlorophyll a concentration of the present invention;

[0052] Figure 3 This is a precision verification diagram of the interpolation result of the present invention;

[0053] Figure 4 This is a comparison chart between the interpolation result of the present invention and the original data. DETAILED DESCRIPTION

[0054] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0055] Reference Figure 1-Figure 4 A method for interpolating chlorophyll a concentration remote sensing data products comprises the following steps:

[0056] Step A: A three-dimensional tensor formed from the original chlorophyll a concentration remote sensing time series to be interpolated Rearranged into a new three-dimensional space-time tensor form according to the characteristics of space, time and period

[0057] In this embodiment, the chlorophyll a concentration remote sensing time series product is a daily chlorophyll a concentration remote sensing image product obtained by the MODIS-Aqua satellite, with a spatial resolution of 4 km and a year range from 2013 to 2022. The specific rearrangement steps are as follows (eg Figure 2 shown):

[0058] Arrange the original chlorophyll a concentration remote sensing time series images into a three-dimensional tensor in chronological order, recorded as:

[0059] Where m and n represent the size of the image, which are 400 rows and 280 columns respectively, and t represents the length of the date in the entire time series, which is 3650 days;

[0060] Since the concentration of chlorophyll a has the characteristic of periodic change, the original tensor is rearranged into a new three-dimensional tensor, recorded as:

[0061] Where ny represents the length of the entire time series, which is 10 years, and nd is the number of observations per year, which is 365 days;

[0062] Step B: Construct an adaptive spatiotemporal tensor completion model based on the tensor truncated nuclear norm model, and input the rearranged chlorophyll a concentration spatiotemporal tensor The tensor completion model is solved by adaptively allocating tensor weights in different dimensions to obtain the spatiotemporal tensor. The solution;

[0063] Step C: Transform the space-time tensor The solution is transformed into the format corresponding to the original data after the rearrangement in step A, and the interpolated chlorophyll a concentration remote sensing time series is obtained.

[0064] In this embodiment, the specific implementation process of constructing the adaptive spatiotemporal tensor completion model is as follows:

[0065] Based on the principle of rank-minimizing tensor completion, the data interpolation problem can be expressed as the following function model:

[0066]

[0067] Where rank is the algebraic rank of the chlorophyll a concentration remote sensing time series data tensor, Ω represents the index set of the observed pixels, and P Ω represents the orthogonal projection on Ω;

[0068] Using the weighted convex nuclear norm to approximate the tensor rank, the above formula is converted to:

[0069]

[0070] In the formula, Representing a tensor The kth expansion matrix of k yes The non-negative weights of ||·|| * represents the nuclear norm regularization term;

[0071] Then, based on low-rank matrix completion and using tensor trace norm instead of matrix nuclear norm, the low-rank tensor completion problem in the previous step is converted to:

[0072]

[0073] Where Y represents the missing tensor, A is a diagonal matrix with diagonal elements of 0 and 1;

[0074] When a pixel in Y is missing, the value is 0, otherwise it is 1;

[0075] is the fidelity term of the data, is a regularization term, which represents the prior distribution of the real data, and λ is a regularization parameter, which is used to balance the contribution of the above two items.

[0076] The specific process of solving the tensor completion model by adaptively allocating tensor weights in different dimensions is as follows:

[0077] Rearranged 3D tensor The k-th dimension expansion matrix is ​​obtained by singular value decomposition method Singular values ​​of and arrange in descending order Construct a one-dimensional singular value sequence;

[0078] Connect the first point in the sequence to the maximum value to form a straight line, subtract each point in the sequence from the straight line, and then normalize the maximum and minimum values ​​to obtain a new singular value difference sequence

[0079] Calculate the low-rank index for each dimension where i k-max is the local maximum of the singular value difference sequence, q k is the length of the singular value sequence in k dimensions;

[0080] Calculate the weight index of different dimensions k of the tensor in

[0081] Furthermore, the space-time tensor is obtained The specific process of the solution is:

[0082] After obtaining the weights of different expansion dimensions k, a new tensor value variable M is introduced k Replace X in step B13 k , and obtain the new model solution formula:

[0083]

[0084] Further define a Lagrangian function:

[0085]

[0086] In the formula, B k is the Lagrangian variable, β is the penalty parameter;

[0087] Based on the above steps, define the optimization sub-problem to be solved by the model:

[0088]

[0089] Solve the subproblem M k The formula for X is:

[0090]

[0091] Set the number of iterations i and update the Lagrangian variable B k In addition, the weight parameter α in each iteration k are updated by solving the tensor completion model. k The update formula is as follows:

[0092]

[0093] In this embodiment, the solution obtained in step B is According to step A, the rearrangement process is reversed to The interpolated chlorophyll a concentration remote sensing time series is obtained.

[0094] The accuracy verification results of the chlorophyll a concentration remote sensing time series after interpolation by the method of the present invention are as follows: Figure 3 The comparison results of the original data and the interpolated data are as follows: Figure 4 It can be seen that, in view of the serious missing problem in the time series of daily chlorophyll a concentration remote sensing data products currently acquired by satellites, the present invention considers the high correlation between the spatial, interannual changes and periodic characteristics of the chlorophyll a concentration remote sensing time series, and performs spatiotemporal rearrangement of the three-dimensional tensor constructed by the original chlorophyll a concentration remote sensing time series to obtain a tensor form with a lower rank characteristic; secondly, a low-rank remote sensing time series tensor completion method with adaptive weight allocation is constructed based on the tensor truncated nuclear norm model; finally, based on the method proposed in the present invention, the missing areas of the daily chlorophyll a concentration data products acquired by satellites are filled. Compared with the existing chlorophyll a concentration remote sensing data interpolation technology, the method of the present application ensures the interpolation accuracy without the need for additional data collection for training, which greatly improves the practical application capability of seriously missing satellite data products in marine environment monitoring.

[0095] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for interpolating chlorophyll a concentration remote sensing data products, characterized by: The following steps are involved: Step A: A three-dimensional tensor formed from the original chlorophyll a concentration remote sensing time series to be interpolated Rearranged into a new three-dimensional space-time tensor form according to the characteristics of space, time and period Step B: Construct an adaptive spatiotemporal tensor completion model based on the tensor truncated nuclear norm model, and input the rearranged chlorophyll a concentration spatiotemporal tensor The tensor completion model is solved by adaptively allocating tensor weights in different dimensions to obtain the spatiotemporal tensor. The solution; Step C: Transform the space-time tensor The solution is transformed into the format corresponding to the original data after the rearrangement in step A, and the interpolated chlorophyll a concentration remote sensing time series is obtained.

2. The method for interpolating chlorophyll a concentration remote sensing data products according to claim 1, characterized in that: The specific steps in step A are as follows: A1. Arrange the original chlorophyll a concentration remote sensing time series images in chronological order into a three-dimensional tensor, recorded as: Where m and n represent the size of the image, and t represents the length of the date in the entire time series; A2. Since the concentration of chlorophyll a has the characteristic of periodic change, the original tensor is rearranged into a new three-dimensional tensor, recorded as: Where ny represents the length of the entire time series in years, and nd is the number of observations per year.

3. The method for interpolating chlorophyll a concentration remote sensing data products according to claim 1, characterized in that: The specific implementation process of constructing the adaptive spatiotemporal tensor completion model is as follows: B11. Functional expression of the rank-minimization tensor completion problem: Where rank is the algebraic rank of the chlorophyll a concentration remote sensing time series data tensor, Ω represents the index set of the observed pixels, and P Ω represents the orthogonal projection on Ω; B12. Using the weighted convex kernel norm to approximate the tensor rank, the above formula is converted to: In the formula, Representing a tensor The kth expansion matrix of k yes The non-negative weights of ||·|| * represents the nuclear norm regularization term; B13. Based on low-rank matrix completion and using tensor trace norm instead of matrix nuclear norm, the low-rank tensor completion problem in the previous step is converted to: Where Y represents the missing tensor, A is a diagonal matrix with diagonal elements of 0 and 1; When a pixel in Y is missing, the value is 0, otherwise it is 1; is the fidelity term of the data, is a regularization term, which represents the prior distribution of the real data, and λ is a regularization parameter, which is used to balance the contribution of the above two items.

4. The method for interpolating chlorophyll a concentration remote sensing data products according to claim 3, characterized in that: The specific process of solving the tensor completion model by adaptively allocating tensor weights in different dimensions is as follows: B21. Rearranged 3D Tensors The k-th dimension expansion matrix is ​​obtained by singular value decomposition method The singular values ​​of Construct a one-dimensional singular value sequence; B22. Connect the first point in the sequence to the maximum value to form a straight line, subtract each point in the sequence from the straight line, and then normalize the maximum and minimum values ​​to obtain a new singular value difference sequence B23. Calculate the low-rank index for each dimension where i k-max is the local maximum of the singular value difference sequence, q k is the length of the singular value sequence in k dimensions; B24. Calculate the weight index of different dimensions k of the tensor in 5. The method for interpolating chlorophyll a concentration remote sensing data products according to claim 4, characterized in that: The space-time tensor is obtained The specific process of the solution is: B31. After obtaining the weights of different expansion dimensions k, introduce a new tensor value variable M k Replace X in step B13 k , and obtain the new model solution formula: Further define a Lagrangian function: In the formula, B k is the Lagrangian variable, β is the penalty parameter; B33. Based on the above steps, define the optimization sub-problem to be solved by the model: B34, solve the sub-problem M k The formula for X is: B35. Set the number of iterations i and update the Lagrangian variable B k In addition, the weight parameter α in each iteration k are updated by solving the tensor completion model. k The update formula is as follows:

6. The method for interpolating chlorophyll a concentration remote sensing data products according to claim 5, characterized in that: The solution obtained in step B According to step A, the rearrangement process is reversed to The interpolated chlorophyll a concentration remote sensing time series is obtained.

Citation Information

Cited By

  • Chronic disease risk assessment and intervention strategy generation system based on data analysis

    CN120221116A