Spatial-temporal characteristic analysis system for earth surface coverage

Abstract

The invention discloses a spatial-temporal characteristic analysis system for earth surface coverage, and relates to the technical field of remote sensing, and the key points of the technical scheme are as follows: the system obtains NDVI time sequence data through a data acquisition module, and removes noise through re-projection, rearrangement and data reconstruction of a data preprocessing module. Afterwards, the NDVI sequence is processed through an intrinsic orthogonal decomposition module, time and space changes are separated, dimension reduction is conducted on feature vectors through a t distribution random neighbor embedding algorithm module, and space distribution data are obtained; and finally, the visual analysis module maps the data and the time coefficient to realize visual analysis of spatial-temporal characteristic change. The system can efficiently capture main modes of land cover changes, provides scientific basis for climate change and land utilization change mode analysis, and has the beneficial effect of improving analysis accuracy and efficiency.

Description

Technical Field

[0001] The present invention relates to the field of remote sensing technology, and more specifically, to a spatiotemporal characteristic analysis system for land cover. Background Art

[0002] Satellite data is the main source of information for monitoring terrestrial ecosystems today. Decades of satellite time series provide a data basis for quantitatively describing global surface vegetation activities and their changing trends. The Normalized Difference Vegetation Index (NDVI) quantitatively analyzes vegetation conditions by measuring the difference between near-infrared (strong reflection by vegetation) and red light (absorption by vegetation), so changes in NDVI are considered to represent changes in vegetation. The spatiotemporal changes in vegetation are often used to assess the productivity and degradation of land caused by natural and human activities. On the one hand, vegetation, as a major component of the terrestrial biosphere, is considered a key factor in climate change. Secondly, the spatiotemporal changes in vegetation are considered to be closely related to land use patterns. Therefore, the analysis of the spatiotemporal variation characteristics of NDVI is of great significance for the analysis of climate change and land use change patterns in a region.

[0003] The study of land cover changes is mainly conducted from two aspects: time and space. The former analyzes the changing patterns of land cover types over time, while the latter generally analyzes from two aspects: horizontal gradient and vertical gradient. Some auxiliary factors are also added to participate in the analysis, such as terrain factors, soil factors, climate factors, economic factors, etc. Based on some vegetation index data (NDVI, EVI, etc.) or Landsat data, the changes in land cover in the study area are analyzed from the core area to the peripheral area and from low altitude to high altitude using buffer analysis, cluster analysis, network analysis, etc. from both quantitative and qualitative perspectives. Exploring the spatiotemporal evolution of land cover can more intuitively and clearly show how the land cover in the region has changed over time from both qualitative and quantitative perspectives.

[0004] Therefore, for the spatiotemporal characteristics of NDVI time series data, a new analysis method is found, which can analyze the spatiotemporal distribution characteristics of the data from a new perspective. Summary of the invention

[0005] The purpose of the present invention is to provide a spatiotemporal feature analysis system for surface cover to solve the problems of data processing complexity, insufficient separation of spatiotemporal features, insufficient visualization and interpretation, low analysis efficiency, and insufficient adaptability to land use changes in the prior art.

[0006] The above technical objectives of the present invention are achieved through the following technical solutions: a spatiotemporal characteristic analysis system for land cover, the system comprising the following modules:

[0007] A data acquisition module is used to acquire normalized vegetation index time series data;

[0008] Data preprocessing module, used to reproject, rearrange and reconstruct NDVI time series data to remove noise;

[0009] The intrinsic orthogonal decomposition module is used to perform intrinsic orthogonal decomposition on the NDVI sequence, separating the time-varying flow field into the sum of the product of a spatial function that does not vary with time and a time function that only depends on time;

[0010] The t-distributed random nearest neighbor embedding algorithm module is used to reduce the dimension of the feature vector obtained by the intrinsic orthogonal decomposition to obtain spatial distribution data;

[0011] The visualization analysis module is used to plot spatial distribution data and time coefficients to analyze changes in spatiotemporal characteristics in a visual way.

[0012] The present invention is further configured as follows: the data preprocessing module specifically includes the following units:

[0013] A reprojection unit, used to convert the sinusoidal projection of the image into a latitude and longitude projection;

[0014] A rearrangement unit is used to rearrange the newly projected NDVI data in time sequence;

[0015] The data reconstruction unit is used to extract the time series of each pixel and reconstruct the data to filter and eliminate the random noise caused by clouds or atmospheric influences, and make the sampling interval of the time dimension of the sequence consistent through reconstruction.

[0016] The present invention is further configured as follows: the intrinsic orthogonal decomposition module specifically includes the following units:

[0017] A covariance matrix construction unit, used to construct a covariance matrix;

[0018] An eigenvalue and eigenvector calculation unit, used to calculate the eigenvalue and eigenvector of the covariance matrix, and obtain the eigenvalue and the corresponding eigenvector;

[0019] The variance contribution rate calculation unit is used to calculate the variance contribution of a certain eigenvalue in the spatial mode and the cumulative variance contribution rate of the previous eigenvector.

[0020] The present invention is further configured as follows: the t-distributed random neighbor embedding algorithm module specifically includes the following units:

[0021] Gaussian distribution similarity calculation unit, used to calculate the similarity distance between any two feature points in high-dimensional space;

[0022] A joint probability calculation unit, used to calculate and normalize the joint probability of high-dimensional space;

[0023] t distribution probability density calculation unit, used to calculate the sample probability density in the low-dimensional subspace;

[0024] The KL divergence optimization unit is used to optimize the cost function C by gradient descent method to minimize the KL divergence and obtain the optimal low-dimensional subspace.

[0025] The present invention is further configured as follows: the visual analysis module specifically includes the following units:

[0026] The spatiotemporal characteristic change mapping unit is used to map the spatial distribution data and the time coefficient to analyze the spatiotemporal characteristic changes in a visual way.

[0027] The present invention is further configured as follows: the system can analyze the spatiotemporal distribution characteristics of NDVI time series data from a new perspective by combining intrinsic orthogonal decomposition with a t-distributed random nearest neighbor embedding algorithm.

[0028] By adopting the above technical solution, the present invention provides a new method for analyzing the spatiotemporal characteristics of surface cover by adopting a technical solution combining proper orthogonal decomposition (POD) and t-distributed stochastic neighbor embedding algorithm (TSNE). The POD method can effectively separate the temporal changes and spatial changes in NDVI time series data, and decompose the complex flow field into the sum of the product of a spatial function that does not change with time and a time function that only depends on time changes. This can not only capture part of the full fluctuation dynamic energy in the flow field, but also express the main temporal and spatial changes with as few modes as possible, thereby objectively and quantitatively reflecting the spatial structure and temporal changes of the flow field. The introduction of the TSNE algorithm further reduces the dimensionality of the feature vector obtained by POD decomposition, so that high-dimensional data can maintain its internal structure and global information in low-dimensional space, greatly improving the interpretability of the data and the accuracy of analysis.

[0029] The present invention can identify the main patterns of land cover change through POD decomposition, and these patterns can characterize the main types of NDVI spatial distribution in the region and can reflect the differences in NDVI fluctuations within the year; the TSNE algorithm enables these high-dimensional feature vectors to be effectively visualized in a low-dimensional space, so that analysts can intuitively understand the local structural feature information of the data. This method combining POD and TSNE not only improves the efficiency of analysis, but also enhances the ability to identify the characteristics of land cover change, providing a powerful tool for the analysis of climate change and land use change patterns.

[0030] In summary, the present invention has the following beneficial effects:

[0031] 1. The present invention analyzes the spatiotemporal characteristics of NDVI time series data by combining proper orthogonal decomposition (POD) and t-distributed stochastic nearest neighbor embedding algorithm (TSNE), which can decompose complex spatiotemporal changes into concise modes, greatly reducing the dimension of the data while retaining the core information of the data. The present invention can efficiently capture the main part of the full fluctuation dynamic energy in the flow field through POD decomposition, while the TSNE algorithm can maintain the internal structure and global information of high-dimensional data in low-dimensional space, making the analysis results more intuitive and accurate. The present invention not only improves the accuracy of land cover change analysis, but also significantly improves the efficiency of analysis, providing strong technical support for research and application in related fields. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 It is a flow chart of the implementation scheme in the embodiment of the present invention;

[0033] Figure 2 It is a schematic diagram of the size of 23 eigenvalues ​​corresponding to POD decomposition in an embodiment of the present invention;

[0034] Figure 3 It is a time coefficient curve graph corresponding to the first three modes of 3NDVI data in an embodiment of the present invention;

[0035] Figure 4 3 is a schematic diagram of the frequency spectrum of the time coefficients corresponding to the first three modes of NDVI data in an embodiment of the present invention. DETAILED DESCRIPTION

[0036] In order to enable those skilled in the art to better understand the solution of the present invention, the technical solution of the present invention will be further described in detail below in conjunction with the embodiments of the present invention and the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0037] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features of the embodiments can be combined with each other. The present invention will be described in detail in conjunction with the embodiments. Figure 1-4 The present invention is described in further detail.

[0038] Embodiment 1: A system for analyzing the spatiotemporal characteristics of land cover, such as Figure 1 As shown, the system includes the following modules:

[0039] A data acquisition module is used to acquire normalized vegetation index time series data;

[0040] The data preprocessing module is used to reproject, rearrange and reconstruct the NDVI time series data to remove noise; the data preprocessing module specifically includes the following units:

[0041] A reprojection unit, used to convert the sinusoidal projection of the image into a latitude and longitude projection;

[0042] A rearrangement unit is used to rearrange the newly projected NDVI data in time sequence;

[0043] A data reconstruction unit is used to extract the time series of each pixel and reconstruct the data to filter and eliminate random noise caused by cloud or atmospheric influence, and to make the sampling interval of the time dimension of the sequence consistent through reconstruction;

[0044] The intrinsic orthogonal decomposition module is used to perform intrinsic orthogonal decomposition processing on the NDVI sequence, and separate the time-varying flow field into the sum of the product of the spatial function that does not vary with time and the time function that only depends on time. The intrinsic orthogonal decomposition module specifically includes the following units:

[0045] A covariance matrix construction unit, used to construct a covariance matrix;

[0046] An eigenvalue and eigenvector calculation unit, used to calculate the eigenvalue and eigenvector of the covariance matrix, and obtain the eigenvalue and the corresponding eigenvector;

[0047] The variance contribution rate calculation unit is used to calculate the variance contribution of a certain eigenvalue in the spatial mode and the cumulative variance contribution rate of the previous eigenvector;

[0048] The t-distributed random neighbor embedding algorithm module is used to perform dimension reduction processing on the feature vector obtained by the intrinsic orthogonal decomposition to obtain spatial distribution data; the t-distributed random neighbor embedding algorithm module specifically includes the following units:

[0049] Gaussian distribution similarity calculation unit, used to calculate the similarity distance between any two feature points in high-dimensional space;

[0050] A joint probability calculation unit, used to calculate and normalize the joint probability of high-dimensional space;

[0051] t distribution probability density calculation unit, used to calculate the sample probability density in the low-dimensional subspace;

[0052] The KL divergence optimization unit is used to optimize the cost function C by gradient descent method to minimize the KL divergence and obtain the optimal low-dimensional subspace.

[0053] The visualization analysis module is used to map the spatial distribution data and the time coefficient to analyze the changes in the spatiotemporal characteristics in a visual way; the visualization analysis module specifically includes the following units: a spatiotemporal characteristic change mapping unit, which is used to map the spatial distribution data and the time coefficient to analyze the changes in the spatiotemporal characteristics in a visual way.

[0054] The system of the present invention combines the Proper orthogonal decomposition (POD) and the t-distributed Stochastic Neighbor Embedding (TSNE) algorithm to analyze the spatiotemporal characteristics of the NDVI time series. The specific flow chart of the scheme is shown in the attached figure. Figure 1 shown.

[0055] First, obtain the NDVI time series data. The specific processing steps are as follows:

[0056] (1) Reproject the NDVI time series data and convert the sinusoidal projection of the image into longitude and latitude projection.

[0057] (2) The newly projected NDVI data were rearranged in chronological order.

[0058] (3) Data reconstruction to remove noise. The time series of each pixel is extracted from the NDVI time series image and the data is reconstructed. The purpose of data reconstruction is twofold: one is to filter and eliminate or weaken the random noise caused by clouds or atmospheric influences; the other is to make the sampling interval of the time dimension of the sequence consistent through reconstruction;

[0059] (4) POD decomposition to obtain eigenvalues ​​and eigenvectors. POD processing is performed on the NDVI series to separate the temporal and spatial variations of the flow field, and to decompose the flow field that varies with time into the sum of the products of a spatial function that does not vary with time and a temporal function that only depends on time.

[0060] The POD method can separate the temporal and spatial variations of the flow field, and decompose the flow field that varies with time into the sum of the products of a spatial function that does not vary with time and a time function that only depends on time. The spatial function is the POD mode. Each mode captures a part of the total fluctuation dynamic energy in the flow field and expresses the main temporal and spatial variations with as few modes as possible, thereby objectively and quantitatively reflecting the spatial structure and temporal variations of the flow field.

[0061] The spatial mode of POD can be obtained through the basis function of eigenanalysis, that is, first construct a covariance matrix, and then calculate its eigenvector, which is the spatial mode of POD. The projection of the covariance matrix to each spatial mode is the time coefficient of the corresponding spatial mode.

[0062] Assume u'(x, t) is a velocity-fluctuating flow field, that is, the time-averaged steady velocity field has been subtracted from the velocity vector X = (x, y, z) represents position, t represents time, POD decomposes the random vector field u'(x, t) into a set of certain function spaces, and the random time coefficient is a k (t), so that

[0063]

[0064] The above formula is considered to be the decomposition result of the fixed spatial mode and the random time coefficient. This is the POD decomposition. The so-called "P" of POD is to find a proper space so that The dynamic energy is maximized through the first n spatial modes. "O" means that these modes are orthonormal, that is, in the appropriate function space,

[0065]

[0066] The orthogonal property is useful because it implies that each time coefficient a k (t) depends only on the spatial mode Φ k (x),

[0067]

[0068] When space is used as a variable and time is used as another variable, POD is essentially symmetric about space and time. It can also be seen as the decomposition of a certain temporal mode through the adjustment of random spatial coefficients. In other words, the positions of space and time are interchangeable.

[0069] The calculation process of POD decomposition is as follows: Assume that the image sequence of NDVI of a certain area is U(x,t), and the image size is n=N x ×N y , there are m=N time points t , then the image sequence minus the mean value forms a matrix

[0070]

[0071] The covariance matrix of matrix U is denoted as C, that is is an m×m matrix. Calculate the eigenvalues ​​and eigenvectors of C and obtain m eigenvalues ​​λ1,λ2,…,λ m , generally arranged from large to small, and the m eigenvectors corresponding to the eigenvalues ​​are recorded as Φ k (x), k = 1, ..., m. If the matrix U is in the matrix Φ (Φ k is its column vector) is denoted as matrix A, then A m×m =Um×n Φ n×m , each column of A corresponds to Φ k The time coefficient, A k =UΦ k .

[0072] When performing POD analysis, it is also necessary to calculate the variance contribution of a certain eigenvalue in the spatial mode, which can describe the detailed spatial distribution of the variance contribution of each mode. The larger the eigenvalue λ is, the more typical the corresponding spatial mode is, and the greater the contribution to the total variance is. That is, the contribution rate of the matrix U field corresponding to the kth eigenvector is:

[0073]

[0074] The cumulative variance contribution rate of the first p eigenvectors is The cumulative variance contribution rate is the basis for retaining the spatial modal order.

[0075] (5) TSNE reduces the dimensionality of the feature vector to obtain spatial distribution data, while preserving the internal structure and global information of the data points.

[0076] The TSNE algorithm aims to discover the nested low-dimensional manifold structure that may exist in the high-dimensional space, and fuse the high-dimensional data to obtain the local structural feature information of the data. Its core idea is to transform the high-dimensional space data X = {x1, x2, ..., x N} to low-dimensional space data Y = {y1, y2, ..., y n} to map, while keeping the distribution probability between each other unchanged to the greatest extent, the low-dimensional space distribution of the data is obtained by minimizing the KL divergence. TSNE first defines the conditional probability P ji , represents any two feature points x that obey Gaussian distribution i and x j The similarity distance between

[0077]

[0078] Where: i The center point is x i Gaussian variance of ||x i -x j || is x i and x j The Euclidean distance between them. Then the conditional probability P j|i Sum and normalize to obtain the high-dimensional space joint probability P with symmetric properties ij for:

[0079]

[0080] Using t distribution to calculate the sample probability density q in low-dimensional subspace ij :

[0081]

[0082] The KL divergence is used to measure the similarity between two distributions, and the cost function C is optimized by the gradient descent method to minimize it to obtain the optimal low-dimensional subspace. The formula is as follows:

[0083]

[0084] Thus, we can obtain low-dimensional data Y = {y1, y2, ..., y n}.

[0085] (6) Plot the spatial distribution data and time coefficients to visualize the changes in spatiotemporal characteristics. The following analysis uses the NDVI data of Ningxia Hui Autonomous Region in 2010 as an example. This includes analysis of spatial distribution characteristics and analysis of temporal distribution characteristics;

[0086] Spatial distribution characteristics analysis: The sizes of the 23 eigenvalues ​​corresponding to POD decomposition are shown in the attached figure. Figure 2 As shown in Table 1, the eigenvalues, variance contribution rates and cumulative variance contribution rates of the first three modes after POD decomposition are shown in Table 1. It can be seen that the cumulative variance contribution rates of the first three modes have exceeded 99.5%, and their eigenvectors can well explain the spatial distribution type of NDVI in the region.

[0087] Table 1 The first three eigenvalues ​​of POD

[0088]

[0089]

[0090] From Table 1 above, we know that the variance contribution rate of the first eigenvalue accounts for 93.9%. Mode 1 is the spatial distribution of the eigenvector corresponding to the first eigenvalue. Mode 1 can characterize the main type of NDVI spatial distribution in the region. The eigenvectors of mode 1 are all negative, indicating that the fluctuation of NDVI in the region is consistent, that is, NDVI shows the characteristics of increasing and decreasing at the same time.

[0091] From Table 1 above, we can see that the variance contribution rate of the second eigenvalue is 18.0305%. Mode 2 corresponds to the spatial distribution of the second eigenvector. The eigenvector of mode 2 can directly reflect the differences in NDVI fluctuations in the region within the year. In addition, from the absolute value of the second-order mode, the areas with large absolute values ​​are mainly distributed in the Yellow River channel area and the Yellow River irrigation area, indicating that there is a small pattern in this area.

[0092] Mode 3 gives the spatial distribution of the corresponding third eigenvector. It can be seen that its eigenvector has more negative values ​​than positive ones. The positive values ​​are mainly concentrated in the plain Yellow River irrigation area and the urban areas along the Yellow River, while the negative values ​​are distributed in most areas, indicating the difference in NDVI oscillation intensity between urban and non-urban areas.

[0093] Analysis of temporal distribution characteristics: The time coefficients corresponding to the first three modes of the NDVI spatiotemporal series in Ningxia Hui Autonomous Region in 2010 are shown in the attached figure. Figure 3 As shown in the figure, the frequency spectra of the time coefficients corresponding to the three modes are shown in the attached figure. Figure 4 As shown, the spectrum is displayed in logarithmic form.

[0094] By the attached Figure 3-4 It can be seen that the three time coefficients are all periodic functions if extended. The time coefficient curve corresponding to the first mode (see Appendix Figure 3 The star line shows that the vegetation change in the area is mainly based on a one-year cycle, which is consistent with the climate characteristics of the region; since the eigenvectors of mode 1 are all negative, from May to September, due to the negative time coefficient, the time coefficient and the spatial mode are multiplied, making the NDVI fluctuation positive, and the region is generally green during this period. The NDVI fluctuation in the region shows a change from weak to strong and then weak again within a year. In addition, since the time coefficient is positive or negative, the larger the absolute value of the coefficient, the more typical the spatial distribution type of the NDVI fluctuation at that moment. The time coefficient curve corresponding to the first mode has the largest absolute value at the 14th time point, that is, mid-July, which is the time when the local vegetation is most prosperous.

[0095] The maximum amplitude of the time coefficient curve corresponding to mode 2 and mode 3 is much smaller than the maximum amplitude of the time coefficient corresponding to mode 1, indicating that mode 2 and mode 3 are relatively weak compared to mode 1. The time coefficient curves corresponding to mode 2 and mode 3 also show periodic characteristics. The time point when the amplitude extreme value of the time coefficient curve of mode 3 occurs is consistent with that of mode 1, indicating that the extreme point is related to the pattern of the vegetation class, but the time point of the amplitude extreme value of the time coefficient curve of mode 2 is slightly different, indicating that these two extreme values ​​correspond to different types of local spatial patterns. The above specific embodiments are merely explanations of the present invention, and they are not limitations of the present invention. After reading this specification, those skilled in the art can make modifications to the present embodiment without creative contribution as needed, but as long as they are within the scope of the claims of the present invention, they are protected by the patent law.

Claims

1. A spatiotemporal characteristic analysis system for land cover, characterized by: The system includes the following modules: A data acquisition module is used to acquire normalized vegetation index time series data; Data preprocessing module, used to reproject, rearrange and reconstruct NDVI time series data to remove noise; The intrinsic orthogonal decomposition module is used to perform intrinsic orthogonal decomposition on the NDVI sequence, separating the time-varying flow field into the sum of the product of a spatial function that does not vary with time and a time function that only depends on time; The t-distributed random nearest neighbor embedding algorithm module is used to perform dimensionality reduction processing on the eigenvectors obtained by the intrinsic orthogonal decomposition to obtain spatial distribution data; The visualization analysis module is used to plot spatial distribution data and time coefficients to analyze changes in spatiotemporal characteristics in a visual way.

2. The system according to claim 1, characterized in that: The data preprocessing module specifically includes the following units: A reprojection unit, used to convert the sinusoidal projection of the image into a latitude and longitude projection; A rearrangement unit is used to rearrange the newly projected NDVI data in time sequence; The data reconstruction unit is used to extract the time series of each pixel and reconstruct the data to filter and eliminate the random noise caused by clouds or atmospheric influences, and make the sampling interval of the time dimension of the sequence consistent through reconstruction.

3. The system according to claim 1, characterized in that: The intrinsic orthogonal decomposition module specifically includes the following units: A covariance matrix construction unit, used to construct a covariance matrix; An eigenvalue and eigenvector calculation unit, used to calculate the eigenvalue and eigenvector of the covariance matrix, and obtain the eigenvalue and the corresponding eigenvector; The variance contribution rate calculation unit is used to calculate the variance contribution of a certain eigenvalue in the spatial mode and the cumulative variance contribution rate of the previous eigenvector.

4. The system according to claim 1, characterized in that: The t-distributed random neighbor embedding algorithm module specifically includes the following units: Gaussian distribution similarity calculation unit, used to calculate the similarity distance between any two feature points in high-dimensional space; A joint probability calculation unit, used to calculate and normalize the joint probability of high-dimensional space; t distribution probability density calculation unit, used to calculate the probability density of samples in low-dimensional subspace; The KL divergence optimization unit is used to optimize the cost function C by gradient descent method to minimize the KL divergence and obtain the optimal low-dimensional subspace.

5. The system according to claim 1, characterized in that: The visual analysis module specifically includes the following units: The spatiotemporal characteristic change mapping unit is used to map the spatial distribution data and the time coefficient to analyze the spatiotemporal characteristic changes in a visual way.

6. The system according to any one of claims 1 to 5, characterized in that: The system can analyze the spatiotemporal distribution characteristics of NDVI time series data from a new perspective by combining intrinsic orthogonal decomposition with the random nearest neighbor embedding algorithm of t distribution.

Images