CN111062267B - Time series remote sensing image dimension reduction method - Google Patents
Time series remote sensing image dimension reduction method Download PDFInfo
- Publication number
- CN111062267B CN111062267B CN201911192701.9A CN201911192701A CN111062267B CN 111062267 B CN111062267 B CN 111062267B CN 201911192701 A CN201911192701 A CN 201911192701A CN 111062267 B CN111062267 B CN 111062267B
- Authority
- CN
- China
- Prior art keywords
- remote sensing
- pixels
- time
- sensing image
- standard sample
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/10—Terrestrial scenes
- G06V20/13—Satellite images
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/22—Matching criteria, e.g. proximity measures
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/23—Clustering techniques
- G06F18/232—Non-hierarchical techniques
- G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
- G06F18/23213—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions with fixed number of clusters, e.g. K-means clustering
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Probability & Statistics with Applications (AREA)
- Astronomy & Astrophysics (AREA)
- Remote Sensing (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The invention provides a time series remote sensing image dimension reduction method, which comprises the following steps: constructing time sequence remote sensing image data, and calculating DTW (dynamic time warping) measurement distance between pixels; selecting a sample to construct a standard sample library; constructing a neighbor graph and giving weight through a kNN algorithm; calculating the geodesic line distance between the pixels of the standard sample library and all the pixels in the time series remote sensing image data; constructing a geodesic distance matrix of all pixels in the pixels of the standard sample library and the time series remote sensing image data and a geodesic distance ion matrix among the pixels of the standard sample library; and obtaining a result after dimensionality reduction by using an MDS algorithm. The method greatly simplifies the preprocessing process of the time series remote sensing image, directly removes invalid data, does not need to perform necessary operations in the traditional method such as interpolation or filtering on the data, avoids the reduction of subsequent processing precision caused by errors caused by the operations, simultaneously obviously reduces algorithm complexity, and can be used for large-area scale remote sensing images.
Description
Technical Field
The invention relates to a remote sensing image processing method, in particular to a time series remote sensing image dimension reduction method.
Background
The term "land cover" refers to a general term for a vegetation cover and an artificial cover on the earth's surface, and is a comprehensive reflection of natural vegetation and elements of the earth's surface covered by natural structures and artificial structures. The land cover is the necessary information for human to know nature and master nature laws, and is the most basic data required by various resource management and geographic information services. Therefore, the acquisition, analysis and updating of land cover information is of great importance.
The remote sensing image data is always an important means for land coverage detection due to the characteristics of macroscopicity and real-time property. The remote sensing data has the advantages that the remote sensing data contains abundant spatial information, and the spatial characteristics of ground features can be researched. But the occurrence of surface events and the evolution of surface features are carried out along with the advance of time, and show a certain change rule in time and space. Therefore, the classification of the remote sensing image cannot depend on the spatial characteristics of the ground features only, and an appropriate method should be adopted to analyze the temporal characteristics of the ground features of the remote sensing image, so as to try to extract the information and knowledge related to the temporal characteristics of the remote sensing image. With the increasing abundance of remote sensing sensor resources, the time resolution of remote sensing data is remarkably improved, and the remote sensing image data accumulated in the same region for a long time contains abundant time, space and spectral information, thereby providing an excellent data source for dynamic monitoring of land coverage. Therefore, establishing a remote sensing image time sequence data set and performing land cover classification based on long-term accumulated remote sensing data becomes a development trend of remote sensing image classification.
Data redundancy exists in the time series remote sensing images, particularly in the area where the land coverage type is not changed. Therefore, before performing land cover classification on the time-series remote sensing images, dimension reduction processing is often performed. Dimension reduction is a method of transforming high-dimensional data into a low-dimensional space while preserving useful information and reducing noise.
The time series remote sensing image is subjected to multi-directional scattering in the imaging process, so that the internal structure of the data presents nonlinear characteristics, and therefore, the nonlinear dimension reduction method is more suitable for processing the time series remote sensing image than the linear dimension reduction method. In nonlinear dimension reduction, the most important method is a dimension reduction method based on manifold learning. The main manifold learning methods include equidistant feature mapping, local linear embedding, local reservation embedding, neighbor reservation embedding and the like. Specific methods can be found in reference 1: balasubramanian M, Schwartz E L. "The isomap algorithm and The cosmetic stability". Science,2002,295(5552):7-7, reference 2: de ladder and r.p.dunn, "localization linear embedding for classification," Pattern Recognition Group, depth.of Imaging Science & Technology, depth University of Technology, Delft, The Netherlands, technology.rep.ph-2002-01, pp.1-12,2002, and document 3: x.he, d.cai, s.yan, and h. -j.zhang, "neighbor preserving embedding," in Computer Vision,2005.ICCV 2005.tent IEEE International Conference on,2005, pp.1208-1213, however, when the above method is directly applied to time-series remote sensing images, there are problems as follows: firstly, the time-series remote sensing images often contain invalid data to a certain extent due to cloud and snow coverage, if the method is adopted, the time-series remote sensing images need to be preprocessed by interpolation or cloud removal, but effective information is not added to the preprocessed images, and a larger prediction error is brought to the contrary, so that the subsequent classification precision is reduced; secondly, when the nonlinear dimension reduction algorithm is used for a large-area time series remote sensing image, the algorithm complexity is high, and the execution is difficult, for example, an equidistant feature mapping algorithm needs to calculate geodesic distances between all pixels, and when the number of image pixels is large, the calculation amount is quite large.
Disclosure of Invention
The invention aims to solve the problem of providing a time series remote sensing image dimension reduction method, which can simplify the preprocessing process of the time series remote sensing image and greatly reduce the complexity of an algorithm.
In order to achieve the above object, the present invention provides a time series remote sensing image dimension reduction method, which comprises the following steps:
s101, constructing time sequence remote sensing image data and calculating a DTW (dynamic time warping) measurement distance between pixels;
s102, selecting a sample to construct a standard sample library;
s103, constructing a neighbor graph and giving weight through a kNN algorithm;
s104, calculating the geodesic line distances between the pixels of the standard sample library and all the pixels in the time series remote sensing image data;
s105, constructing a geodesic distance matrix of all pixels in the pixels of the standard sample library and the time series remote sensing image data and a geodesic distance ion matrix among the pixels of the standard sample library;
and S106, obtaining a result after dimensionality reduction by using an MDS algorithm.
Preferably, the step of step S101 is as follows:
1a) arranging the original time series remote sensing images according to a spectrum-time phase;
1b) eliminating invalid data and keeping valid data;
1c) and calculating DTW measurement distance between pixels in the time sequence remote sensing image data by DTW measurement.
Further preferably, the arrangement method of the original time series remote sensing images in the step 1a) comprises: all the phase data of the same band are arranged together.
Further preferably, in the step 1b), the cloud mask file is used to mask the data of the cloud-covered area of all the wave bands in each time phase from the constructed time series remote sensing image data.
Preferably, the step of step S102 is as follows:
2a) respectively selecting a field sample point for each category on the time sequence, and extracting a corresponding spectrum-time phase curve;
2b) integrating the effective data of all samples in each category to form a complete spectrum-time phase standard sample curve in each category;
2c) according to the standard sample curve, searching a plurality of pixels with the minimum distance from the time series remote sensing image according to DTW measurement to construct a standard sample library;
2d) determining a clustering center of the pixel in the step 2c) by adopting a K mean value clustering method;
2e) and taking the pixel closest to the clustering center as the final standard sample library pixel.
Specifically, the step S103 is as follows:
3a) selecting k pixels with the minimum distance from each pixel on the time sequence remote sensing image according to DTW measurement, and connecting the pixels by edges;
3b) each edge is given a weight, i.e. DTW measures distance.
Preferably, in the step S104, a Dijkstra algorithm is adopted to calculate geodesic distances between the pixels in the standard sample library and all the pixels on the time series remote sensing image.
Preferably, the step of step S106 is as follows: and taking the distance sub-matrix constructed in the step S105 as input, searching an optimal solution in space by using an MDS algorithm, and performing affine linear transformation on the geodesic distance matrix based on the optimal solution to obtain a dimension reduction result of the whole image.
Through the technical scheme, the invention has the following beneficial effects:
(1) the invention greatly simplifies the preprocessing process of the time series remote sensing image, directly removes invalid data, does not need to carry out necessary operations in the traditional methods such as interpolation or filtering on the data, and avoids the reduction of subsequent processing precision caused by errors caused by the operations;
(2) the Euclidean distance of pixel similarity calculated in the traditional dimension reduction method is replaced by DTW measurement, and the DTW measurement can measure data with unequal dimensions, so that the DTW measurement improved dimension reduction method can be better suitable for time sequence remote sensing images;
(3) by establishing a standard sample library and then only calculating the similarity between the pixels of the standard sample library and all data points instead of calculating the similarity between all data points, the complexity of the algorithm is greatly reduced, the method can be used for large-area time series remote sensing images, and the operation is beneficial to improving the classification accuracy of subsequent time series remote sensing images;
(4) compared with the traditional mode of only arranging phases in time, the method has the advantages that the spectrum-time phase data arrangement mode is adopted, the characteristic of the ground object and the climate is more prominent, and the improvement of the classification precision of the ground surface coverage is facilitated.
Drawings
FIG. 1 is a flowchart of a time series remote sensing image dimension reduction method according to an embodiment of the present invention;
FIG. 2 is an image of a region of interest in an embodiment of the present invention;
FIG. 3 is a spectrum-time phase sequence standard sample curve of corn, rice, and soybean in an example of the present invention;
FIG. 4 is a schematic diagram of time series remote sensing images of any 2 wave bands in 161 original wave bands before dimension reduction in the embodiment of the present invention;
fig. 5 is a schematic diagram of a dimension reduction result of 2 wave bands of the time-series remote sensing image in the embodiment of the present invention.
Detailed Description
The following detailed description of specific embodiments of the present invention is provided in connection with the accompanying drawings and examples. It should be understood that the detailed description and specific examples, while indicating the present invention, are given by way of illustration and explanation only, not limitation.
The invention relates to Landsat8 (Landsat 8 is the eighth satellite of the United states terrestrial satellite program (Landsat) in 2015 23 Ind of Shangyang City of Liaoning province, and the Landsat is successfully carried and launched by an Atlas-V rocket in No. 2 and 11 in 2013 at the air force base of California Van denburg, and is originally called as a 'terrestrial satellite Data Continuity Mission' (LDCM). The Land Imager (OLI) and the Thermal Infrared Sensor (TIRS) remote sensing image carried by the Landsat8 are taken as examples (as shown in fig. 2, the size of the image of the research area is 7921 × 8041, 63692761 pixels are totally calculated, if the matrix is calculated according to the existing method, 63692761 × 63692761 is needed, and the calculated amount is very huge), and the specific implementation mode of the time series remote sensing image dimension reduction method is explained. The shooting times of the 23 experimental images were 2015 year 1 month 5 days, 2015 year 1 month 21 days, 2015 year 2 month 6 days, 2015 year 2 month 22 days, 2015 year 3 month 10 days, 2015 year 3 month 26 days, 2015 year 4 month 11 days, 2015 year 4 month 27 days, 2015 year 5 month 13 days, 2015 year 5 month 29 days, 2015 year 6 month 14 days, 2015 year 6 month 30 days, 2015 year 7 month 16 days, 2015 year 8 month 1 days, 2015 year 8 month 17 days, 2015 year 9 month 2 days, 2015 year 9 month 18 days, 2015 year 10 month 4 days, 2015 year 10 month 20 days, 2015 year 11 month 5 days, 2015 year 11 month 21 days, 2015 year 12 month 7 days, 2015 year 12 month 23 days. Each image uses wave bands 1-7 and a cloud mask file, image data is downloaded in an official website (http:// earth x plorer. usgs. gov /), CFmask (cloud coverage evaluation) files contained in the data are made into the cloud mask file, and the cloud mask file establishing process is as follows: 1) using Landsat 8OLI L1T (L1T: level 1Terrain, Level 1 Terrain) reflectivity product, which contains a CFmask file detected by a CFmask algorithm, wherein pixel values in the CFmask only include 0, 1, 2, 3, 4 and 255, and respectively represent the meanings of pure water, shadow, snow, cloud and filling; 2) and reserving the pixels with the pixel values of 0 and 1 in the CFmask file, and establishing a cloud mask file.
Referring to fig. 1, the time series remote sensing image dimension reduction method according to the embodiment of the invention includes the following steps:
s101, constructing Time sequence remote sensing image data, and calculating a Dynamic Time Warping (DTW) measurement distance between elements;
in this embodiment, the Landsat8 image 1-7 waveband time-series remote sensing image data of the above 23 time phases are arranged together according to the order of the 1 st waveband all time phases, the 2 nd waveband all time phases, the 3 rd waveband all time phases, the 4 th waveband all time phases, the 5 th waveband all time phases, the 6 th waveband all time phases and the 7 th waveband all time phases, so as to construct time-series remote sensing image data including 161 wave bands. Then, masking the data of the cloud-covered area of all the wave bands in each time phase from the constructed time series remote sensing image data by adopting the established cloud mask file, so that all effective data (namely pixels corresponding to pixels with pixel values of 0 and 1 in the cloud mask file in the image) are reserved, and invalid data are directly removed. And finally, calculating the distances among all the pixels according to a calculation formula of DTW measurement, wherein the calculation formula is as follows:
assuming that the time sequence length of the pixel M is t and the time sequence length of the pixel N is s, then
M＝m1，m2，…，mi，…，mt
N＝n1，n2，…，nj，…，ns
The DTW distance of M and N is:
γ(i，j)＝d(i，j)+min[γ(i-1，j-1)，γ(i1，j)，γ(i，j-1)]
wherein d (i, j) represents miAnd njThe distance between them is calculated as follows:
d(i，j)＝(ti-sj)2
the calculation process needs to satisfy the following constraints:
(1) monotonicity: the phase must be monotonically increasing, i.e. mi≤mi+1And nj≤nj+1；
(2) The characteristics of the bounding: the calculation is performed from the first point (m)1，n1) Start to last point (m)t，ns) Finishing;
(3) continuity: when calculating the distance, each time one step is carried out, a near point, namely m, must be connectedi+1-mi1 or less and nj+1-nj≤1。
The traditional dimension reduction method requires that the original data dimensions are required to be consistent, and for the time series remote sensing image, as the image is easily covered by snow and cloud in the imaging process, invalid data appears in the image, if the traditional method is adopted, the data needs to be subjected to interpolation or filtering supplementary processing, and the processing does not increase data information, but brings larger prediction error, so that the subsequent processing precision of the image is realized. The invention adopts DTW measurement to replace the Euclidean distance for calculating the pixel similarity in the traditional dimension reduction method, and the DTW measurement can measure data with unequal dimensions, so the dimension reduction method improved by adopting the DTW measurement can be better suitable for time sequence remote sensing images.
S102, selecting a sample to construct a standard sample library;
in this embodiment, the remote sensing image mainly includes three crops of corn, rice and soybean, 20 samples are collected for each crop on the spot, time series remote sensing image data corresponding to each sample is respectively extracted, time series curves of 20 samples of each type are integrated, and it is ensured that each time phase of each wavelength band has data, so that each crop forms a complete spectrum-time phase series curve (as shown in fig. 3). Then, 1000 pixels each having the smallest distance from its complete spectrum-phase sequence curve are selected for each crop based on the DTW metric. And then clustering the 3000 pixels based on a K-means method, wherein the clustering method comprises the following steps: initially, a standard sample curve (as shown in fig. 3) is selected as a cluster center, 3000 pixels are allocated to a set corresponding to the cluster center closest to the pixel center, and then the cluster center of each set is updated to an average value of all points in the set, that is, the cluster center is updated to an average value of all points in the set
Wherein u isiIs the center of the cluster, and,
C＝{C1，C2，…，Ckis a collection of clusters.
And continuing to perform iterative operation, and finishing iteration when the square error of the clustering division is minimum. And taking the 300 image elements closest to the clustering center as a standard sample library of the category.
The traditional dimension reduction method, especially the global dimension reduction method, such as Isomap (Isometric feature mapping) algorithm, needs to calculate the similarity between all data points, which is difficult to be executed for large-area time sequence remote sensing images due to excessive pixel number.
For time series remote sensing images, the characteristic of the feature of.
S103, constructing a Neighbor graph and giving weight through a K-Nearest Neighbor classification algorithm (kNN);
in this embodiment, the number k of neighbors in kNN is set to 10, 10 pixels with the smallest DTW metric distance in each pixel (the standard sample library pixel and all image pixels) are connected by edges to form a neighbor map, and then the weights of the edges are set to the DTW distances between two corresponding pixels respectively (that is, the value corresponding to this position is written as the DTW distance in the matrix, and if not, the value corresponding to the nearest 10 pixels is assigned as 0 in the matrix).
S104, calculating the geodesic line distances between the pixels of the standard sample library and all the pixels in the time series remote sensing image data;
in this embodiment, the standard sample library has 900 pixels, and the geodesic distances between the 900 pixels and all pixels in the time-series remote sensing image data are calculated based on Dijkstra algorithm (Dijkstra algorithm), respectively, and the calculation process is as follows:
(1) regarding each pixel on the standard sample library and the time series remote sensing image as a node on the graph, selecting a node x in the standard sample library as a starting point set A, and enabling other nodes to belong to another set B;
(2) calculating the distance d [ x ] from other nodes to the starting point (if adjacent, d [ x ] is the side weight, if not, d [ x ] is infinity);
(3) selecting the minimum d [ x ], adding the node y corresponding to the d [ x ] into the set A, and removing the node y from the set B;
(4) and updating the distance d [ y ] between the nodes adjacent to the y, wherein the distance d [ y ] is min { d [ y ], d [ x ] + d [ x, y }, and d [ x, y ] represents the distance from the node x to the nodes adjacent to the y.
(5) Repeating the steps 3 and 4 until the target point z is added into the set A, wherein the d [ z ] corresponding to the target point z is the shortest path length, namely the geodesic distance between the node x and the node z.
S105, constructing a distance matrix;
in this embodiment, a geodesic distance matrix of 900 pixels and all pixels (this matrix is 900 rows by all pixel number columns) and a geodesic distance ion matrix, that is, a geodesic distance matrix between pixels in the standard sample library (900 rows by 900 columns) are established according to the geodesic distances calculated in step S104.
And S106, obtaining a result after dimensionality reduction by using an MDS (Multidimensional Scaling) algorithm.
In this embodiment, the geodesic distance ion matrix obtained in step S105 is used as an input, an MDS algorithm is used to find a global optimal solution in space (a solving process of the MDS algorithm is to perform eigenvalue decomposition on a constructed geodesic distance matrix to obtain eigenvalues and eigenvectors, the eigenvector corresponding to the largest eigenvalue is the found optimal solution), then an affine linear transformation is performed on the geodesic distance matrix (i.e., the distance matrix of the above 900 pixels and all pixels) based on the optimal solution to obtain a dimension reduction result of the whole image), and a calculating process is as follows:
assuming that the geodesic distance matrix is D, the geodesic distance ion matrix, that is, the geodesic matrix between pixels of the standard sample library is Ds, and the pixels of the standard sample library are reduced to D-dimensional sample matrix Zs, we aim to obtain the matrix Z from image pixels reduced to D-dimensional:
let the distance between the ith pixel and the jth pixel in Ds be dist [ i, j]In Zs is | | Zi-ZjI (matrix i row minus 1 norm after j row), and dist [ i, j]＝||Zi-Zj||
The distance matrix B after the dimension reduction of the geodesic distance matrix D is equal to ZTZ, then
Thereby obtaining:
for ease of discussion, let the sample matrix Zs be centered, i.e.:
it is possible to obtain:
order:
(2) (3) is brought into (1) to obtain
Performing eigenvalue decomposition on the distance matrix B subjected to dimensionality reduction, wherein B is V lambada VTThen the Zs expression can be obtained:
and finally, obtaining a result Z after dimension reduction of the time series remote sensing image by using affine linear transformation:
wherein m is the number of pixels, ZTThe inversion of the Z is carried out,
is the pseudo-inverse of the Zs,
is the column mean of the geodesic distance matrix D.
After the distance matrix B is decomposed, eigenvectors corresponding to the first d eigenvalues are usually reserved (d may be any value, generally, information is reserved in data with eigenvalues greater than 0, and the larger the eigenvalue is, the more information is included, in this embodiment, d is the number of data with eigenvalues greater than 0, and is 10). The dimension reduction effect is shown in fig. 5 (each point on the graph represents 1 pixel, 1 pixel has 10 bands, that is, 10 values, and these 10 values can be drawn as coordinates of 10 dimensions, but such multiple dimensions cannot be visually displayed, and for convenience of display, 2 dimensions of 10 are selected in fig. 5), compared with fig. 4 before dimension reduction (because the correlation of time series remote sensing image data is strong, any two bands show similar results, and the horizontal and vertical coordinates in fig. 4 are both enlarged by 10000 times), after dimension reduction by using the method of the present invention, the discrimination of three types of crops is more obvious, more information is expressed by using less data, and the subsequent classification processing is facilitated.
The method greatly simplifies the preprocessing process of the time series remote sensing image, directly removes invalid data, does not need to perform necessary operations in the traditional methods such as interpolation or filtering on the data, and avoids the reduction of subsequent processing precision caused by errors caused by the operations; the Euclidean distance of pixel similarity calculated in the traditional dimension reduction method is replaced by DTW measurement, and the DTW measurement can measure data with unequal dimensions, so that the DTW measurement improved dimension reduction method can be better suitable for time sequence remote sensing images; by establishing a standard sample library and then only calculating the similarity between the pixels of the standard sample library and all data points instead of calculating the similarity between all data points, the complexity of the algorithm is greatly reduced (in the embodiment, only a matrix of 900 × 63692761 is required to be calculated, and compared with a matrix of 63692761 × 63692761, the calculated amount is greatly reduced), the method can be used for the remote sensing images of the large-area time sequence, and the operation is helpful for improving the classification accuracy of the subsequent remote sensing images of the time sequence; compared with the traditional mode of only arranging time phases, the method has the advantages that the spectrum-time phase data arrangement mode is adopted, the characteristic of the ground object climate is more prominent, and the improvement of the classification precision of the ground surface coverage is facilitated.
The preferred embodiments of the present invention have been described in detail with reference to the accompanying drawings and examples, however, the present invention is not limited to the specific details of the above embodiments, and various simple modifications can be made to the technical solution of the present invention within the technical idea of the present invention, and these simple modifications are within the protective scope of the present invention.
It should be noted that the various technical features described in the above embodiments can be combined in any suitable manner without contradiction, and the invention is not described in any way for the possible combinations in order to avoid unnecessary repetition.
In addition, any combination of the various embodiments of the present invention is also possible, and the same should be considered as the disclosure of the present invention as long as it does not depart from the spirit of the present invention.
Claims (6)
1. A time series remote sensing image dimension reduction method is characterized by comprising the following steps:
s101, constructing time sequence remote sensing image data and calculating a DTW (dynamic time warping) measurement distance between pixels;
s102, selecting a sample to construct a standard sample library;
s103, constructing a neighbor graph and giving weight through a kNN algorithm;
s104, calculating the geodesic distance between the pixels of the standard sample library and all the pixels in the time series remote sensing image data based on the neighbor map and the weight obtained in the step S103;
s105, constructing a geodesic distance matrix of all pixels in the pixels of the standard sample library and the time sequence remote sensing image data and a geodesic distance ion matrix among the pixels of the standard sample library according to the geodesic distance obtained in the step S104;
s106, obtaining a result after dimensionality reduction by using an MDS algorithm;
wherein, the step of step S103 is as follows:
3a) selecting k pixels with the minimum distance from each pixel on the time sequence remote sensing image according to DTW measurement, and connecting the pixels by edges to form a neighbor map;
3b) giving a weight to each edge, wherein the weight is the DTW measurement distance between two corresponding pixels;
the step of step S106 is as follows: and taking the geodesic distance ion matrix constructed in the step S105 as input, searching an optimal solution in space by using an MDS algorithm, and performing affine linear transformation on the geodesic distance matrix based on the optimal solution to obtain a dimension reduction result of the whole image.
2. The dimension reduction method for the time-series remote sensing images according to claim 1, wherein the step S101 is as follows:
1a) arranging the original time series remote sensing images according to a spectrum-time phase;
1b) eliminating invalid data and keeping valid data;
1c) and calculating DTW measurement distance between pixels in the time sequence remote sensing image data by DTW measurement.
3. The dimension reduction method for the time-series remote sensing images according to claim 2, wherein the arrangement method for the original time-series remote sensing images in the step 1a) comprises the following steps: all phase data of the same band are arranged together.
4. The dimension reduction method for the time series remote sensing images according to claim 2, wherein in the step 1b), the cloud mask file is used for masking the data of the cloud covered area of all the wave bands in each time phase from the constructed time series remote sensing image data.
5. The dimension reduction method for the time-series remote sensing images according to claim 1, wherein the step S102 comprises the following steps:
2a) respectively selecting a field sample point for each category on the time sequence, and extracting a corresponding spectrum-time phase curve;
2b) integrating the effective data of all samples in each category to form a complete spectrum-time phase standard sample curve in each category;
2c) according to the standard sample curve, searching a plurality of pixels with the minimum distance from the time series remote sensing image according to DTW measurement to construct a standard sample library;
2d) determining the clustering center of the pixel in the step 2c) by adopting a K mean value clustering method;
2e) and taking the pixel closest to the clustering center as the final standard sample library pixel.
6. The dimension reduction method for the time-series remote sensing images according to claim 1, wherein step S104 adopts Dijkstra algorithm to calculate the geodesic distance between the pixels in the standard sample library and all the pixels on the time-series remote sensing images.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201911192701.9A CN111062267B (en) | 2019-11-28 | 2019-11-28 | Time series remote sensing image dimension reduction method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201911192701.9A CN111062267B (en) | 2019-11-28 | 2019-11-28 | Time series remote sensing image dimension reduction method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN111062267A CN111062267A (en) | 2020-04-24 |
| CN111062267B true CN111062267B (en) | 2022-07-08 |
Family
ID=70299243
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201911192701.9A Active CN111062267B (en) | 2019-11-28 | 2019-11-28 | Time series remote sensing image dimension reduction method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN111062267B (en) |
Families Citing this family (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112070130B (en) * | 2020-08-25 | 2023-03-28 | 西安交通大学 | Cross-road vehicle matching method and system based on dynamic time warping |
| CN113125351B (en) * | 2021-03-25 | 2022-11-29 | 国家卫星气象中心(国家空间天气监测预警中心) | Multi-time remote sensing image optimization synthesis method and system |
| CN117138590A (en) * | 2023-10-30 | 2023-12-01 | 南通睿智超临界科技发展有限公司 | Method for detecting treatment effect of forward osmosis membrane |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102043857A (en) * | 2010-12-27 | 2011-05-04 | 中国科学院计算技术研究所 | All-nearest-neighbor query method and system |
| CN102521843A (en) * | 2011-11-28 | 2012-06-27 | 大连大学 | Three-dimensional human body motion analysis and synthesis method based on manifold learning |
| CN105719323A (en) * | 2016-02-18 | 2016-06-29 | 江苏优图空间信息科技有限公司 | Hyperspectral dimension reducing method based on map optimizing theory |
| EP3059682A1 (en) * | 2015-02-17 | 2016-08-24 | Schneider Electric Industries SAS | Data-processing and modelling system for analysing the energy consumption of a site |
-
2019
- 2019-11-28 CN CN201911192701.9A patent/CN111062267B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102043857A (en) * | 2010-12-27 | 2011-05-04 | 中国科学院计算技术研究所 | All-nearest-neighbor query method and system |
| CN102521843A (en) * | 2011-11-28 | 2012-06-27 | 大连大学 | Three-dimensional human body motion analysis and synthesis method based on manifold learning |
| EP3059682A1 (en) * | 2015-02-17 | 2016-08-24 | Schneider Electric Industries SAS | Data-processing and modelling system for analysing the energy consumption of a site |
| CN105719323A (en) * | 2016-02-18 | 2016-06-29 | 江苏优图空间信息科技有限公司 | Hyperspectral dimension reducing method based on map optimizing theory |
Non-Patent Citations (6)
| Title |
|---|
| Improved time series land cover classification by missing-observation-adaptive nonlinear dimensionality reduction;L. Yan 等;《Remote Sensing of Environment》;20141212;第158卷;478-491 * |
| Land Cover Classification Using Integrated Spectral,Temporal, and Spatial Features Derived from Remotely Sensed Images;Yongguang Zhai 等;《remote sensing》;20180301;1-26 * |
| The Isomap Algorithm and Topological Stability;Mukund Balasubramanian 等;《SCIENCE》;20020104;第295卷;7a * |
| 基于流形学习的人体动作识别方法研究;蒋玮;《中国优秀博硕士学位论文全文数据库(硕士) 信息科技辑》;20170315(第3期);I138-5204 * |
| 基于非线性降维时序遥感影像的作物分类;翟涌光 等;《农业工程学报》;20181031;第34卷(第19期);177-183 * |
| 综合时空谱多维遥感特征的作物精细分类方法研究——以东北三省为例;翟涌光;《中国优秀博硕士学位论文全文数据库(博士) 农业科技辑》;20171115(第11期);D043-3 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN111062267A (en) | 2020-04-24 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111062267B (en) | Time series remote sensing image dimension reduction method | |
| CN109919875B (en) | High-time-frequency remote sensing image feature-assisted residential area extraction and classification method | |
| CN110120046B (en) | Potential landslide identification method integrating DEM (digital elevation model), optical remote sensing and deformation information | |
| US20060233421A1 (en) | Methods and apparatus for adaptive foreground background analysis | |
| CN112381013B (en) | Urban vegetation inversion method and system based on high-resolution remote sensing image | |
| CN111652193A (en) | Wetland classification method based on multi-source images | |
| CN107194936B (en) | Hyperspectral image target detection method based on superpixel combined sparse representation | |
| CN107274441B (en) | Wave band calibration method and system for hyperspectral image | |
| Lavanya et al. | Terrain mapping of LandSat8 images using MNF and classifying soil properties using ensemble modelling | |
| AU2021102192A4 (en) | Remote sensing classification and identification method, device, equipment and storage medium for soil types | |
| CN115236655A (en) | Landslide identification method, system, equipment and medium based on fully-polarized SAR | |
| Karakacan Kuzucu et al. | Testing the potential of vegetation indices for land use/cover classification using high resolution data | |
| CN114662626A (en) | Lithology automatic classification method and system for multi-source remote sensing data multi-feature fusion | |
| CN115372282A (en) | Farmland soil water content monitoring method based on hyperspectral image of unmanned aerial vehicle | |
| CN115240072A (en) | Hyperspectral multi-class change detection method based on multidirectional multi-scale spectrum-space residual convolution neural network | |
| Bektas Balcik et al. | Determination of land cover/land use using spot 7 data with supervised classification methods | |
| CN113850769B (en) | Hyperspectral change detection method based on Simese space spectrum joint convolution network | |
| Orlíková et al. | Land cover classification using sentinel-1 SAR data | |
| CN115144357A (en) | Mineral identification method based on airborne mid-infrared hyperspectral remote sensing data of mining area | |
| CN113446998B (en) | Hyperspectral target detection data-based dynamic unmixing method | |
| Cresson et al. | Comparison of convolutional neural networks for cloudy optical images reconstruction from single or multitemporal joint SAR and optical images | |
| CN107358625B (en) | SAR image change detection method based on SPP Net and region-of-interest detection | |
| Agmalaro et al. | Sentinel 1 classification for garlic land identification using Support Vector Machine | |
| Jusoff et al. | Mapping of individual oil palm trees using airborne hyperspectral sensing: an overview | |
| Hall | Remote sensing applications in viticulture: recent advances and new opportunities |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |