Reconstruction method and system for vegetation coefficient

Abstract

The invention belongs to the technical field of vegetation coefficients, and discloses a reconstruction method and system for vegetation coefficients, and the method comprises the steps: obtaining a vegetation coefficient array of a satellite image as input, and representing the array as three-dimensional tensor data; rejecting impervious surface data in the tensor, and dividing the data in the tensor into a low type, a middle type and a high type according to the vegetation coverage; for different types of data, average values of tensors are calculated along different dimensions, spaces and times of the tensors respectively, then the average values are subtracted from initial data, so that the data are centered along the space and time axes, and unknown data values are initialized; calculating a reconstruction error, namely a normalized root-mean-square error NRMSE, of the iteration, if the error is smaller than a required value, ending the iteration and outputting a reconstruction tensor, otherwise, carrying out iteration again; and adding the reconstructed tensor to the subtracted average value to complete inverse centering, and finally combining different types of reconstructed tensors to obtain a reconstructed vegetation index. According to the method, the missing vegetation index is iteratively reconstructed by using the tensor decomposition algorithm for different types of data, so that the effectiveness and accuracy of reconstruction are ensured.

Description

Technical Field

[0001] The invention belongs to the technical field of vegetation coefficients, and in particular relates to a method and system for reconstructing vegetation coefficients. Background Art

[0002] Vegetation index is an indicator used in remote sensing to evaluate vegetation conditions. It is usually calculated based on the reflectance or radiation characteristics of vegetation in different spectral bands. Common vegetation indices such as the Normalized Difference Vegetation Index (NDVI), the Ratio Vegetation Index (RVI), and the Difference Vegetation Index (DVI) are all calculated based on the reflectance of the near-infrared band and the reflectance of the red light band. Vegetation indices are widely used and can be used for vegetation growth status monitoring, vegetation coverage assessment, and crop health status analysis. However, the acquisition of vegetation indices depends on remote sensing images, so when the satellite sensor fails or captures a large area of ​​cloud coverage, the vegetation index cannot be calculated. In order to solve the problem of missing remote sensing data, many methods for reconstructing or predicting vegetation coefficients have emerged.

[0003] Since vegetation index reconstruction can be expressed as a spatiotemporal data recovery problem, the mainstream methods are mainly divided into: time series analysis, data mining, machine learning algorithms, and deep learning algorithms. The time series analysis method uses historical vegetation index data to predict future trends through time series models (such as ARIMA models); data mining methods use techniques such as decision trees and cluster analysis to extract patterns and trends of vegetation indices from large amounts of remote sensing data; machine learning algorithms use various machine learning techniques, such as support vector machines (SVMs), random forests, and gradient boosting machines (GBMs), to establish prediction models; deep learning algorithms use convolutional neural networks (CNNs), long short-term memory networks (LSTMs), or a combination of them to learn the spatiotemporal characteristics of vegetation indices from remote sensing images. In short, the principles of these methods are to extract the spatiotemporal characteristics of historical vegetation index data to reconstruct future vegetation indices.

[0004] Therefore, how to effectively extract the temporal and spatial features in vegetation index data and how to combine temporal and spatial features will determine the quality of prediction performance. A large number of methods simply mine the temporal dependency of data while ignoring the spatial dependency of data. Although some methods comprehensively consider temporal and spatial features, they do not achieve good results in combining different features. Summary of the invention

[0005] In view of the problems existing in the prior art, the present invention provides a method and system for reconstructing vegetation coefficients.

[0006] The present invention is achieved by a reconstruction method for vegetation coefficients, the method comprising:

[0007] S1. Get the vegetation coefficient array of the satellite image as input, and the array is represented as a three-dimensional tensor data;

[0008] S2, remove the impervious surface data in the tensor, and divide the data in the tensor into three types: low, medium and high according to the vegetation coverage;

[0009] S3. For different types of data, calculate the average value of the tensor along different dimensions, space and time of the tensor, and then subtract these average values ​​from the initial data to center the data along the space and time axes and initialize the unknown data values;

[0010] S4. For different types of data, the tensor decomposition algorithm PARAFAC is used to calculate the tensor decomposition of the current tensor representing the data in this iteration, estimate the current reconstructed data based on the tensor decomposition, and fill in the unknown values ​​in the reconstructed data;

[0011] S5. Calculate the reconstruction error of this iteration, i.e., the normalized root mean square error NRMSE. If the error is less than the required value, terminate the iteration and output the reconstructed tensor. Otherwise, repeat the iteration.

[0012] S6. The reconstructed tensor is added with the subtracted mean value to complete the decentering, and finally the reconstructed tensors of different types are combined to obtain the reconstructed vegetation index.

[0013] Further, the S1 includes:

[0014] The vegetation index contained in each satellite image is a two-dimensional data, which only has the characteristics of spatial dimension, namely latitude and longitude. Multiple satellite images collected over a continuous period of time are represented as a three-dimensional tensor data, whose dimensions are latitude, longitude and time. Therefore, the three-dimensional tensor includes the characteristics of both spatial and temporal dimensions.

[0015] Further, the S2 includes:

[0016] Auxiliary data are used to remove impervious surface data such as urban roads, reservoirs, and rivers in the image tensor. At the same time, the data in the tensor are divided into three types: low, medium, and high according to vegetation coverage.

[0017] Further, the S3 includes:

[0018] For different types of data, the missing data values ​​are ignored, and the spatial dimension average and the time dimension average are calculated along the spatial dimension, latitude and longitude, and time dimension respectively. The corresponding average is subtracted from each data in the tensor, and each data is centered on the spatial axis and time axis. The missing data values ​​are filled with zero values ​​to complete the initialization.

[0019] Further, the S4 comprises:

[0020] For different types of data, the tensor decomposition algorithm PARAFAC is used to calculate the tensor decomposition of the current tensor representing the data, and the reconstructed data is obtained by using the tensor decomposition, and the unknown values ​​in the reconstructed data are filled according to the current tensor data.

[0021] Further, the S5 comprises:

[0022] The reconstruction error of this iteration, namely the normalized root mean square error NRMSE, is calculated. If the error is less than the required value, the iteration is terminated and the reconstructed tensor is output, otherwise the iteration is repeated.

[0023] The normalized root mean square error NRMSE is:

[0024]

[0025] Where y′ represents the missing value reconstructed in the current iteration, y represents the missing value reconstructed in the previous iteration, and σ[y] represents the variance of the reconstructed missing value in the previous iteration.

[0026] Further, the S6 comprises:

[0027] If the reconstruction error meets the requirements, the centered reconstruction tensor is output, and then the de-centering is performed, that is, each data in the reconstruction tensor is added with the average value in the spatial and temporal dimensions to maintain the consistency of the data before and after reconstruction. Finally, different types of reconstruction data are merged to obtain the reconstructed vegetation index.

[0028] Further, the PARAFAC tensor decomposition algorithm and solution process are as follows:

[0029] Suppose the input three-dimensional tensor is Where I represents the latitude, J represents the longitude, and K represents the number of time periods.

[0030] The goal of the PARAFAC algorithm is to iteratively calculate the best approximation of χ using the alternating least squares ALS algorithm, assuming that the number of rank-one components, that is, the number of reconstructed vegetation coefficient components R, is fixed. The goals are:

[0031]

[0032] in represents the vector outer product, A, B, C represent the rank-one decomposition matrix of the tensor in different modes or dimensions, λ represents the weight vector, a r , b r 、c r Column vectors representing the r-th rank-one components of different decomposition matrices.

[0033] The ALS algorithm first initializes the decomposition matrix, fixes B and C, and calculates the best A; then fixes A and C, calculates the best B; finally fixes A and B, and calculates the best C. Repeat the above steps until convergence. For example, when B and C are fixed, the problem is reduced to linear least squares, and the objective is rewritten as:

[0034]

[0035] where ⊙ denotes the Khatri–Rao product, T denotes the transpose of the matrix, X (1) represents the component of X under mode 1, and the optimal solution is:

[0036]

[0037] in represents the pseudo-inverse of the matrix. Finally, Perform column normalization to obtain A, that is, for r = 1, ..., R, let Both The calculation of B and C is the same. Repeat the above calculation until convergence, output λ, A, B, C, and have:

[0038]

[0039] Finally, the unknown values ​​in the reconstructed data are filled in according to the current tensor decomposition results, and the reconstruction error is calculated.

[0040] Another object of the present invention is to provide a vegetation coefficient reconstruction system based on the vegetation coefficient reconstruction method, the system specifically comprising:

[0041] An input acquisition module is used to obtain the vegetation coefficient array of the satellite image as input, and the array is represented as a three-dimensional tensor data;

[0042] The elimination module is connected to the input acquisition module to eliminate the impervious surface data in the tensor and classify the data in the tensor into three types: low, medium and high according to the vegetation coverage;

[0043] The calculation module is connected to the elimination module. For different types of data, it calculates the average value of the tensor along different dimensions of the tensor, space and time, and then subtracts these average values ​​from the initial data to center the data along the space and time axes and initialize the unknown data values;

[0044] The tensor decomposition module is connected to the computing module. For different types of data, the tensor decomposition algorithm PARAFAC is used to calculate the tensor decomposition of the current tensor representing the data in this iteration, estimate the current reconstructed data based on the tensor decomposition, and fill in the unknown values ​​in the reconstructed data.

[0045] The normalization module is connected to the tensor decomposition module to calculate the reconstruction error of this iteration, namely the normalized root mean square error NRMSE. If the error is less than the required value, the iteration is terminated and the reconstructed tensor is output, otherwise the iteration is repeated;

[0046] The merging module is connected to the normalization module, and the reconstructed tensor is added with the subtracted mean value to complete the de-centering. Finally, the reconstructed tensors of different types are merged to obtain the reconstructed vegetation index.

[0047] Another object of the present invention is to provide a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the vegetation coefficient reconstruction method.

[0048] In combination with the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solutions to be protected by the present invention are as follows:

[0049] The present invention comprehensively considers the spatial and temporal characteristics of the data, represents the data as a three-dimensional tensor, innovatively classifies the data according to vegetation coverage, and iteratively reconstructs the missing vegetation index for different types of data using a tensor decomposition algorithm, thereby ensuring the effectiveness and accuracy of the reconstruction.

[0050] The technical problems solved by the technical solution of the present invention in industrial applications are as follows:

[0051] 1) Solve the problem of incomplete vegetation coefficient data:

[0052] In the prior art, when using satellite remote sensing images to obtain vegetation coefficient data, some spatial or temporal data are missing due to cloud obstruction, equipment failure, transmission interference, data collection cycle limitations, etc., resulting in incomplete data and affecting the accuracy of vegetation coverage analysis.

[0053] 2) Solve the problem of data noise and outliers:

[0054] Traditional vegetation index reconstruction methods often fail to exclude impervious surface data or fail to classify according to vegetation coverage, resulting in large data noise and high outliers in the reconstruction results, making the reconstruction results unreliable.

[0055] 3) Solve the problem of insufficient accuracy and efficiency of reconstruction algorithm:

[0056] In the existing technology, some reconstruction methods (such as linear interpolation and simple mean filling) are difficult to ensure the consistency and integrity of data in time and space dimensions, and the reconstruction accuracy is low. In addition, some complex algorithms have large computational complexity, slow iterative convergence, and low computational efficiency, making them difficult to apply in large-scale data reconstruction.

[0057] The significant technical progress obtained by the present invention:

[0058] 1) Improve data integrity and accuracy:

[0059] The present invention eliminates impervious surface data and classifies the data into low, medium and high vegetation coverage, thereby effectively eliminating irrelevant data and reducing data noise.

[0060] The tensor decomposition algorithm (PARAFAC) is used to reconstruct and fill in different types of vegetation coefficient data in time and space dimensions, significantly improving the integrity and accuracy of the data.

[0061] 2) Optimize spatial and temporal consistency:

[0062] By calculating the average value along the time axis and the space axis and centering the data, the data is balanced along the two dimensions before and after reconstruction, eliminating data deviation.

[0063] After tensor decomposition and reconstruction, the average value is added back to complete the de-centering, ensuring the coherence and consistency of the data in spatial and temporal dimensions.

[0064] 3) Iterative optimization improves reconstruction accuracy:

[0065] The present invention monitors the reconstruction error (normalized root mean square error NRMSE) in real time through iterative calculation, and the iteration is terminated only when the error is less than a preset threshold.

[0066] This adaptive iterative optimization mechanism greatly improves the reconstruction accuracy and ensures that the final result meets high-precision requirements.

[0067] 4) Enhance computing efficiency and big data adaptability:

[0068] This method adopts the PARAFAC tensor decomposition algorithm, which can efficiently process large-scale three-dimensional tensor data with low computational complexity and is suitable for vegetation index reconstruction over a large range and for a long time.

[0069] Classification of processed data (low, medium, and high vegetation cover) further reduces computational redundancy and improves the efficiency of data reconstruction.

[0070] 5) Extensive industrial application value:

[0071] The present invention can be applied to the fields of ecological monitoring, land use analysis, vegetation restoration assessment, etc., and has significant advantages in particular in cloud-prone areas or urban greening data analysis.

[0072] Through high-precision data reconstruction, accurate and reliable data support can be provided for vegetation coverage analysis, environmental protection policy formulation, agricultural resource planning, etc.

[0073] The present invention significantly improves the integrity, accuracy and efficiency of vegetation coefficient data reconstruction through classification and noise elimination, centering processing, tensor decomposition and iterative optimization, solves the problems of data missing, noise interference and high computational complexity in the prior art, and has broad industrial application value and significant technological progress. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] Figure 1 is a flow chart of a method for reconstructing vegetation coefficients provided by an embodiment of the present invention;

[0075] Figure 2 is a structural diagram of a vegetation coefficient reconstruction system provided by an embodiment of the present invention;

[0076] In the figure: 1. Input acquisition module; 2. Elimination module; 3. Calculation module; 4. Tensor decomposition module; 5. Normalization module; 6. Merge module. DETAILED DESCRIPTION

[0077] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0078] The following are two specific embodiments of the reconstruction method for vegetation coefficients, which elaborate on the application process and effects in different scenarios:

[0079] Example 1: Reconstruction of regional vegetation index based on remote sensing data

[0080] A certain area is affected by long-term uneven rainfall and human activities, resulting in incomplete or abnormal vegetation index data. The three-dimensional tensor data of vegetation coefficients in this area is obtained through satellite image acquisition, and the data needs to be reconstructed to improve data integrity and accuracy.

[0081] 1) Get input data (S1):

[0082] Use satellite remote sensing images (such as MODIS or Landsat) to obtain vegetation index data for the area over a period of time to form a three-dimensional tensor with dimensions including: time, spatial longitude, and spatial latitude.

[0083] 2) Eliminate impervious surface data and classify (S2):

[0084] Eliminate data on impervious surfaces such as buildings and roads.

[0085] According to the vegetation coverage, the regional data are divided into three types: low vegetation coverage (such as 0-30%), medium vegetation coverage (such as 30-70%) and high vegetation coverage (70-100%).

[0086] 3) Data centering (S3):

[0087] The average values ​​of the three types of data are calculated along the time and space dimensions respectively.

[0088] Subtract the corresponding mean from the initial data to center the data and initialize the missing values ​​in the data.

[0089] 4) Tensor decomposition and reconstruction (S4):

[0090] The PARAFAC tensor decomposition algorithm is used to decompose the three types of tensor data: low, medium and high.

[0091] The data were reconstructed based on the decomposition results, missing values ​​were filled, and preliminary reconstructed vegetation index data were obtained.

[0092] 5) Iteration error detection (S5):

[0093] The normalized root mean square error (NRMSE) is calculated to determine whether the reconstruction result meets the preset accuracy requirements (such as NRMSE < 0.05).

[0094] If it is not satisfied, continue iterating the calculation until the error meets the requirement.

[0095] 6) Data de-centering and merging (S6):

[0096] Decentering was accomplished by adding the reconstructed data back to the previously subtracted mean.

[0097] The reconstructed tensor data of three types, low, medium and high, are merged to obtain the complete vegetation index data.

[0098] The reconstruction of vegetation index data was completed in this area, missing values ​​were filled, and the spatial continuity and temporal consistency of the data were significantly improved.

[0099] The reconstruction error NRMSE is 0.03, which meets the accuracy requirements.

[0100] Example 2: Reconstruction and analysis of urban green vegetation index

[0101] When a city was conducting greening assessment, the vegetation index data for some time periods were missing due to cloud cover and equipment interference. This method was used to reconstruct the vegetation coefficient data of the urban greening area to accurately assess the vegetation coverage.

[0102] 1) Get input data (S1):

[0103] Vegetation index data are obtained through remote sensing satellite images of urban areas to form three-dimensional tensor data of time, space (urban grid coordinates) and green area type.

[0104] 2) Eliminate impervious surface data and classify (S2):

[0105] Impervious surface data such as buildings, roads and water bodies are eliminated.

[0106] The data are divided into three categories according to greening areas: park green space (high vegetation coverage), residential greening (medium vegetation coverage) and industrial greening (low vegetation coverage).

[0107] 3) Data centering (S3):

[0108] The average values ​​of the three regional types of data are calculated along the time and space dimensions respectively.

[0109] The mean was subtracted from the data, the data were centered along the time and space axes, and missing values ​​were initialized.

[0110] 4) Tensor decomposition and reconstruction (S4):

[0111] The PARAFAC tensor decomposition algorithm is used on three types of data: park green space, residential greening and industrial greening, to obtain the tensor decomposition representation of each area.

[0112] Estimate data based on tensor decomposition results and fill missing values ​​in a 3D tensor.

[0113] 5) Iteration error detection (S5):

[0114] Calculate the normalized root mean square error (NRMSE) for each iteration.

[0115] When NRMSE is less than 0.04, the iteration ends and the reconstructed data is output.

[0116] 6) Data de-centering and merging (S6):

[0117] The three types of reconstructed data were added back to their respective means.

[0118] The reconstructed park green space, residential greening and industrial area greening data are merged to obtain complete vegetation index data for urban areas.

[0119] The reconstruction of the vegetation index in the urban area was completed, solving the problem of data missing caused by cloud cover.

[0120] The NRMSE of the reconstruction result is 0.035, ensuring high accuracy.

[0121] It provides complete urban greening assessment data to support greening management and planning decisions.

[0122] The above two embodiments demonstrate the application effects of this method in different scenarios, including regional vegetation index reconstruction and urban greening assessment. Both achieve high-precision data reconstruction through tensor decomposition and iterative optimization, and have strong applicability and practical value.

[0123] like Figure 1 As shown, an embodiment of the present invention provides a method for reconstructing vegetation coefficients, the method comprising:

[0124] S1. Get the vegetation coefficient array of the satellite image as input, and the array is represented as a three-dimensional tensor data;

[0125] S2, remove the impervious surface data in the tensor, and divide the data in the tensor into three types: low, medium and high according to the vegetation coverage;

[0126] S3. For different types of data, calculate the average value of the tensor along different dimensions of the tensor, space and time, and then subtract these average values ​​from the initial data to center the data along the space and time axes and initialize the unknown data values;

[0127] S4. For different types of data, the tensor decomposition algorithm PARAFAC is used to calculate the tensor decomposition of the current tensor representing the data in this iteration, estimate the current reconstructed data based on the tensor decomposition, and fill in the unknown values ​​in the reconstructed data;

[0128] S5. Calculate the reconstruction error of this iteration, i.e., the normalized root mean square error NRMSE. If the error is less than the required value, terminate the iteration and output the reconstructed tensor. Otherwise, repeat the iteration.

[0129] S6. The reconstructed tensor is added with the subtracted mean value to complete the decentering, and finally the reconstructed tensors of different types are combined to obtain the reconstructed vegetation index.

[0130] Said S1 comprises:

[0131] The vegetation index contained in each satellite image is a two-dimensional data, which only has the characteristics of spatial dimension, namely latitude and longitude. Multiple satellite images collected over a continuous period of time are represented as a three-dimensional tensor data, whose dimensions are latitude, longitude and time. Therefore, the three-dimensional tensor includes the characteristics of both spatial and temporal dimensions.

[0132] The S2 comprises:

[0133] Auxiliary data are used to remove impervious surface data such as urban roads, reservoirs, and rivers in the image tensor. At the same time, the data in the tensor are divided into three types: low, medium, and high according to vegetation coverage.

[0134] The S3 includes:

[0135] For different types of data, the missing data values ​​are ignored, and the spatial dimension average and the time dimension average are calculated along the spatial dimension, latitude and longitude, and time dimension respectively. The corresponding average is subtracted from each data in the tensor, and each data is centered on the spatial axis and time axis. The missing data values ​​are filled with zero values ​​to complete the initialization.

[0136] The S4 comprises:

[0137] For different types of data, the tensor decomposition algorithm PARAFAC is used to calculate the tensor decomposition of the current tensor representing the data, and the reconstructed data is obtained by using the tensor decomposition, and the unknown values ​​in the reconstructed data are filled according to the current tensor data.

[0138] The S5 comprises:

[0139] The reconstruction error of this iteration, namely the normalized root mean square error NRMSE, is calculated. If the error is less than the required value, the iteration is terminated and the reconstructed tensor is output, otherwise the iteration is repeated.

[0140] The normalized root mean square error NRMSE is:

[0141]

[0142] Where y′ represents the missing value reconstructed in the current iteration, y represents the missing value reconstructed in the previous iteration, and σ[y] represents the variance of the reconstructed missing value in the previous iteration.

[0143] The S6 comprises:

[0144] If the reconstruction error meets the requirements, the centered reconstruction tensor is output, and then the de-centering is performed, that is, each data in the reconstruction tensor is added with the average value in the spatial and temporal dimensions to maintain the consistency of the data before and after reconstruction. Finally, different types of reconstruction data are merged to obtain the reconstructed vegetation index.

[0145] The PARAFAC tensor decomposition algorithm and solution process are as follows:

[0146] Suppose the input three-dimensional tensor is Where I represents the latitude, J represents the longitude, and K represents the number of time periods.

[0147] The goal of the PARAFAC algorithm is to iteratively calculate the number of rank-one components, i.e., the number of reconstructed vegetation coefficient components R, using the alternating least squares ALS algorithm. The best approximation of

[0148] The goals are:

[0149]

[0150] in represents the vector outer product, A, B, C represent the rank-one decomposition matrix of the tensor in different modes or dimensions, λ represents the weight vector, a r , b r 、c r Column vectors representing the r-th rank-one components of different decomposition matrices.

[0151] The ALS algorithm first initializes the decomposition matrix, fixes B and C, and calculates the best A; then fixes A and C, calculates the best B; finally fixes A and B, and calculates the best C. Repeat the above steps until convergence. For example, when B and C are fixed, the problem is reduced to linear least squares, and the objective is rewritten as:

[0152]

[0153] where ⊙ denotes the Khatri–Rao product, T denotes the transpose of the matrix, X represents the component of X in mode 1, and the optimal solution is:

[0154]

[0155] in represents the pseudo-inverse of the matrix. Finally, Perform column normalization to obtain A, that is, for r = 1, ..., R, let Both The calculation of B and C is the same. Repeat the above calculation until convergence, output λ, A, B, C, and have:

[0156]

[0157] Finally, the unknown values ​​in the reconstructed data are filled in according to the current tensor decomposition results, and the reconstruction error is calculated.

[0158] like Figure 2 As shown, an embodiment of the present invention provides a vegetation coefficient reconstruction system based on the vegetation coefficient reconstruction method, and the system specifically includes:

[0159] Input acquisition module 1 is used to obtain the vegetation coefficient array of the satellite image as input, and the array is represented as a three-dimensional tensor data;

[0160] The elimination module 2 is connected to the input acquisition module 1, eliminates the impervious surface data in the tensor, and divides the data in the tensor into three types: low, medium and high according to the vegetation coverage;

[0161] The calculation module 3 is connected to the elimination module 2. For different types of data, the average value of the tensor is calculated along different dimensions of the tensor, space and time, and then these average values ​​are subtracted from the initial data to center the data along the space and time axis and initialize the unknown data value;

[0162] The tensor decomposition module 4 is connected to the calculation module 3. For different types of data, the tensor decomposition algorithm PARAFAC is used to calculate the tensor decomposition of the current tensor representing the data in this iteration, estimate the current reconstructed data based on the tensor decomposition, and fill in the unknown values ​​in the reconstructed data;

[0163] Normalization module 5 is connected to tensor decomposition module 4 to calculate the reconstruction error of this iteration, i.e., normalized root mean square error NRMSE. If the error is less than the required value, the iteration is terminated and the reconstructed tensor is output, otherwise, the iteration is repeated;

[0164] The merging module 6 is connected to the normalization module 5, and the reconstructed tensor is added with the subtracted average value to complete the anti-centering, and finally the reconstructed tensors of different types are merged to obtain the reconstructed vegetation index.

[0165] An embodiment of the present invention provides a computer device, which includes a memory and a processor. The memory stores a computer program. When the computer program is executed by the processor, the processor executes the steps of the vegetation coefficient reconstruction method.

[0166] It should be noted that the embodiments of the present invention can be implemented by hardware, software, or a combination of software and hardware. The hardware part can be implemented using dedicated logic; the software part can be stored in a memory and executed by an appropriate instruction execution system, such as a microprocessor or dedicated design hardware. It can be understood by a person of ordinary skill in the art that the above-mentioned devices and methods can be implemented using computer executable instructions and / or contained in a processor control code, such as a carrier medium such as a disk, CD or DVD-ROM, a programmable memory such as a read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. Such code is provided on the carrier medium. The device and its modules of the present invention can be implemented by hardware circuits such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, etc., or programmable hardware devices such as field programmable gate arrays, programmable logic devices, etc., can also be implemented by software executed by various types of processors, and can also be implemented by a combination of the above-mentioned hardware circuits and software, such as firmware.

[0167] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any modification, equivalent substitution and improvement made by any technician familiar with the technical field within the technical scope disclosed by the present invention and within the spirit and principle of the present invention should be covered within the protection scope of the present invention.

Claims

1. A method for reconstructing vegetation coefficients, characterized in that: The method includes: S1. Get the vegetation coefficient array of the satellite image as input, and the array is represented as a three-dimensional tensor data; S2, remove the impervious surface data in the tensor, and divide the data in the tensor into three types: low, medium and high according to the vegetation coverage; S3. For different types of data, calculate the average value of the tensor along different dimensions of the tensor, space and time, and then subtract these average values ​​from the initial data to center the data along the space and time axes and initialize the unknown data values; S4. For different types of data, the tensor decomposition algorithm PARAFAC is used to calculate the tensor decomposition of the current tensor representing the data in this iteration, estimate the current reconstructed data based on the tensor decomposition, and fill in the unknown values ​​in the reconstructed data; S5. Calculate the reconstruction error of this iteration, i.e., the normalized root mean square error NRMSE. If the error is less than the required value, terminate the iteration and output the reconstructed tensor. Otherwise, repeat the iteration. S6. The reconstructed tensor is added with the subtracted mean value to complete the decentering, and finally the reconstructed tensors of different types are combined to obtain the reconstructed vegetation index.

2. The vegetation coefficient reconstruction method according to claim 1, characterized in that: Said S1 comprises: The vegetation index contained in each satellite image is a two-dimensional data, which only has the characteristics of spatial dimension, namely latitude and longitude. Multiple satellite images collected over a continuous period of time are represented as a three-dimensional tensor data, whose dimensions are latitude, longitude and time. Therefore, the three-dimensional tensor includes the characteristics of both spatial and temporal dimensions.

3. The vegetation coefficient reconstruction method according to claim 1, characterized in that: The S2 comprises: Auxiliary data are used to remove impervious surface data such as urban roads, reservoirs, and rivers in the image tensor. At the same time, the data in the tensor are divided into three types: low, medium, and high according to vegetation coverage.

4. The vegetation coefficient reconstruction method according to claim 1, characterized in that: The S3 includes: For different types of data, the missing data values ​​are ignored, and the spatial dimension average and the time dimension average are calculated along the spatial dimension, latitude and longitude, and time dimension respectively. The corresponding average is subtracted from each data in the tensor, and each data is centered on the spatial axis and time axis. The missing data values ​​are filled with zero values ​​to complete the initialization.

5. The vegetation coefficient reconstruction method according to claim 1, characterized in that: The S4 comprises: For different types of data, the tensor decomposition algorithm PARAFAC is used to calculate the tensor decomposition of the current tensor representing the data, and the reconstructed data is obtained by using the tensor decomposition, and the unknown values ​​in the reconstructed data are filled according to the current tensor data.

6. The vegetation coefficient reconstruction method according to claim 1, characterized in that: The S5 comprises: The reconstruction error of this iteration, namely the normalized root mean square error NRMSE, is calculated. If the error is less than the required value, the iteration is terminated and the reconstructed tensor is output, otherwise the iteration is repeated. The normalized root mean square error NRMSE is: Where y′ represents the missing value reconstructed in the current iteration, y represents the missing value reconstructed in the previous iteration, and σ[y] represents the variance of the reconstructed missing value in the previous iteration.

7. The vegetation coefficient reconstruction method according to claim 1, characterized in that: The S6 comprises: If the reconstruction error meets the requirements, the centered reconstruction tensor is output, and then the de-centering is performed, that is, each data in the reconstruction tensor is added with the average value in the spatial and temporal dimensions to maintain the consistency of the data before and after reconstruction. Finally, different types of reconstruction data are merged to obtain the reconstructed vegetation index.

8. The vegetation coefficient reconstruction method according to claim 5, characterized in that: The PARAFAC tensor decomposition algorithm and solution process are as follows: Assume the input three-dimensional tensor is Where I represents the latitude, J represents the longitude, and K represents the number of time periods. The goal of the PARAFAC algorithm is to iteratively calculate the number of rank-one components, i.e., the number of reconstructed vegetation coefficient components R, using the alternating least squares ALS algorithm. The best approximation of The goals are: in represents the vector outer product, A, B, C represent the rank-one decomposition matrix of the tensor in different modes or dimensions, λ represents the weight vector, a r , b r 、c r Column vectors representing the r-th rank-1 components of different decomposition matrices. The ALS algorithm first initializes the decomposition matrix, fixes B and C, and calculates the best A; then fixes A and C, calculates the best B; finally fixes A and B, and calculates the best C. Repeat the above steps until convergence. For example, when B and C are fixed, the problem is reduced to linear least squares, and the objective is rewritten as: where ⊙ denotes the Khatri–Rao product, T denotes the transpose of the matrix, X (1) represents the component of X under mode 1, and the optimal solution is: in represents the pseudo-inverse of the matrix. Finally, Perform column normalization to obtain A, that is, for r = 1, ..., R, let Both The calculation of B and C is the same. Repeat the above calculation until convergence, output λ, A, B, C, and have: Finally, the unknown values ​​in the reconstructed data are filled in according to the current tensor decomposition results, and the reconstruction error is calculated.

9. A vegetation coefficient reconstruction system based on the vegetation coefficient reconstruction method according to claims 1-8, characterized in that: The system specifically includes: An input acquisition module is used to obtain the vegetation coefficient array of the satellite image as input, and the array is represented as a three-dimensional tensor data; The elimination module is connected to the input acquisition module to eliminate the impervious surface data in the tensor and classify the data in the tensor into three types: low, medium and high according to the vegetation coverage; The calculation module is connected to the elimination module. For different types of data, it calculates the average value of the tensor along different dimensions of the tensor, space and time, and then subtracts these average values ​​from the initial data to center the data along the space and time axes and initialize the unknown data values; The tensor decomposition module is connected to the computing module. For different types of data, the tensor decomposition algorithm PARAFAC is used to calculate the tensor decomposition of the current tensor representing the data in this iteration, estimate the current reconstructed data based on the tensor decomposition, and fill in the unknown values ​​in the reconstructed data. The normalization module is connected to the tensor decomposition module to calculate the reconstruction error of this iteration, namely the normalized root mean square error NRMSE. If the error is less than the required value, the iteration is terminated and the reconstructed tensor is output, otherwise the iteration is repeated; The merging module is connected to the normalization module, and the reconstructed tensor is added with the subtracted mean value to complete the de-centering. Finally, the reconstructed tensors of different types are merged to obtain the reconstructed vegetation index.

10. A computer device, characterized in that: The computer device includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the vegetation coefficient reconstruction method according to any one of claims 1 to 8.

Images