A method for extracting spatio-temporal evolution patterns from geo-temporal sequences

Abstract

The application discloses a method for extracting spatio-temporal evolution patterns from geographical time series, comprising the following steps: obtaining geographical time series data collected by a sensor, time-slicing the geographical time series data to obtain a plurality of time slices; based on each time slice, generating a time window by using a wrapping window strategy, calculating the correlation of any two geographical time series in each time window, and screening out a plurality of geographical time series pairs with strong correlation by using a first threshold; based on the geographical time series pairs with strong correlation, screening out geographical time series pairs with close spatial distance by using a second threshold, and constructing a graph relationship network; based on the graph relationship network, extracting evolution patterns therein by using a community discovery algorithm. The above method provides an evolution pattern extraction method capable of accurately reflecting spatio-temporal relationship.

Description

Technical Field

[0001] This invention relates to the field of time series processing, and more specifically to a method for extracting spatiotemporal evolution patterns from geographic time series. Background Technology

[0002] Today, humans have deployed a large number of sensors across geospatial areas to continuously monitor various spatiotemporal phenomena, such as traffic conditions, air quality, rainfall, and temperature. These geographic sensors operate continuously for extended periods, generating massive amounts of geographic time-series data.

[0003] In most cases, the geographic time series generated by these sensors are not isolated in space and time. When spatiotemporal phenomena develop and evolve within a geographic region, the geographic time series of adjacent geographic sensors within that region will show a correlated trend. For example, when a factory emits air pollutants, the geographic time series of nearby air quality sensors will show a correlated upward trend. An evolutionary pattern is defined as the correlation between spatially adjacent (sensor) geographic time series over a period of time in the temporal dimension.

[0004] In an evolutionary model, any two geographic time series should satisfy the following conditions: 1) Temporal correlation: they should exhibit a correlated trend within the time period of the evolutionary model; 2) Geographical proximity: only time series recorded by spatially adjacent sensors have a meaningful relationship. The temporal correlation between any two geographic time series changes dynamically over time, and this dynamic change is difficult to capture. Furthermore, the temporal correlation and geographic proximity of geographic time series are heterogeneous in terms of measurement scale and semantics, making it difficult enough to fuse them to capture spatiotemporal relationships; uncovering evolutionary patterns among multiple geographic time series is even more challenging.

[0005] Association analysis and clustering methods for general time series (SRAghabozorgi, ASHirkhorshidi, and YWTeh. Time-series clustering: A decade review. Information System, 53:16–38, 2015.) can be used to capture temporal correlations. However, these methods often directly consider the correlation of time series over the entire time span, failing to capture changes in local temporal correlations. Furthermore, they do not consider the geospatial relationships of time series, thus making it difficult to capture the aforementioned evolutionary patterns.

[0006] The co-clustering proposed in references 1 (Wu X, Zurita-Milla R, Kraak M J. Co-clustering geo-referencedtime series: exploring spatio-temporal patterns in Dutch temperature data[J]. International Journal of Geographical Information Science, 2015, 29(4): 624-642.) and 2 (Ullah S, Daud H, Dass SC, et al. Detecting space-time disease clusters with arbitrary shapes and sizes using aco-clustering approach[J]. Geospatial health, 2017, 12(2.)) can take into account both spatial and temporal information and capture local features.

[0007] Specifically, collaborative clustering methods cluster time series based on both spatial distribution and temporal trends. However, time series recorded by sensors geographically close together are not always correlated, while sensors geographically distant may record time series with temporal correlations due to long-lasting and wide-ranging spatiotemporal events. This method processes time series independently in both time and space, thus failing to capture the dynamic evolution patterns within them. Summary of the Invention

[0008] In view of the above, the purpose of this invention is to provide a method for extracting spatiotemporal evolution patterns from geographic time series, so as to solve the problem that current methods cannot capture dynamic evolution patterns in spatiotemporal time, and realize that the extracted evolution patterns can capture implicit spatiotemporal relationships.

[0009] To achieve the above-mentioned objectives, the method for extracting spatiotemporal evolution patterns from geographic time series provided in the embodiments includes the following steps:

[0010] Step 1: Obtain geographic time series data collected by sensors, slice the geographic time series data into multiple time slices, and perform steps 2 to 4 for each time slice;

[0011] Step 2: Based on each time slice, generate a time window using the package window strategy, calculate the correlation between any two geographic time series within each time window, and use the first threshold to filter out multiple geographic time series pairs with strong correlation; the strong correlation refers to the correlation coefficient value being in the range of [0.5, 1).

[0012] Step 3: Based on strongly correlated geographic time series pairs, use a second threshold to filter out geographic time series pairs that are spatially close, and construct a graph relationship network;

[0013] Step 4: Based on the graph relationship network, use the community detection algorithm to extract the evolution patterns.

[0014] Preferably, in step 1, the geographic time series data is sliced ​​into multiple time slices, including: [the process involves] slicing the geographic time series data into time slices. The time span is uniformly divided into T time slices of equal size s. i s i ∈S, S={s1,s2,…s T}, where time slice s i This is represented by a data structure with a start timestamp and an end timestamp, where S represents the set of time slices and i represents the time slice index.

[0015] Typically, geographic time series spans a long period (in years, months, etc.), while individual evolutionary patterns are relatively short (in minutes, hours, days, etc.). To extract local features from geographic time series, it is necessary to first convert the continuous time dimension into discrete time slices. The specific size of the time slice depends on the application scenario; if the time slice size is too small, it will be difficult to capture significant features, while if it is too large, it may be difficult to capture local details and dynamics.

[0016] Preferably, in step 2, generating a time window using a wrap-around window strategy includes: for the i-th time slice s i Generate time window w i At that time, time slice s i As the central time slice, based on the central time slice s i and two adjacent time slices s i-1 ,s i+1 Generate time window w i =(s i-1 .t s ,s i+1 .t e ), where s i-1 .t s This indicates the time slice s preceding the center time slice. i-1 The start timestamp, s i+1 .t eIndicates the next time slice s after the center time slice i+1 The end timestamp.

[0017] Preferably, in step 2, calculating the correlation between any two geographic time series within each time window includes: calculating the correlation of the i-th time window w i Any two geographic time series V a and V b Pearson correlation coefficient r(V) a V b w i ).

[0018] The Pearson correlation coefficient is the most commonly used measure of correlation. However, subtle sudden events or data measurement errors can lead to instability in the correlation coefficient. For example, when using the Pearson correlation coefficient to measure the correlation between two geographic time series over three consecutive time slices, the two geographic time series may be determined to be correlated in the first and last two time slices, but uncorrelated in the middle time slice due to measurement errors of individual values. This results in two changes in the correlation of geographic time series within a very small time span (three time slices): correlated-uncorrelated-correlated. To capture significant and accurate patterns, it is necessary to smooth the temporal correlation of geographic time series. Therefore, this invention proposes a wrapping window strategy to avoid such abrupt changes.

[0019] Preferably, in step 2, a first threshold is used to filter out multiple strongly correlated geographical time series pairs (V). a V b ), including: introducing a first threshold th r Filter out sets of highly correlated geographic time series pairs r(V a V b w i ) represents the i-th time window w i Inland geographic time series pairs (V a V b The correlation coefficient of ) This represents geographic time series data.

[0020] Further, in step 3, the geographical time series pairs that are spatially close are selected using a second threshold, including: using the second threshold th d Filter out the set of highly correlated geographic time series pairs CT wi A set P of geographically close time series pairs wi ={(V a V b )|dis(l a , l b )<th d, (V a V b )∈CT wi}, where dis(l a , l b )<th d Represents the geographic time series V a and geographic time series V b The distance between them is less than the second threshold th d .

[0021] Furthermore, in step 3, a graph relationship network is constructed, including: based on the geographic time series of the set P. wi Construct a graph relation network, which is represented as an unweighted and undirected graph G. wi ={VX wi ,P wi}, where VX wi For P wi The set of all geographic time series contained in a geographic time series, i.e., the set of nodes in the graph, P wi The set of edges in the graph is represented in the form of geographic time series pairs. The unweighted, undirected edge between two nodes in the graph represents the geographic time series V. a V b In time slices i They are internally related and spatially adjacent.

[0022] Preferably, in step 4, the community detection algorithm uses the Louvain algorithm.

[0023] Furthermore, the Louvain algorithm is used to extract the community to which each geographic time series belongs, i.e., the evolutionary pattern, wherein each geographic time series can only belong to one community.

[0024] Compared with the prior art, the beneficial effects of the present invention include at least the following:

[0025] 1) Spatiotemporal Feature Fusion Relationship Network Modeling: This invention preserves the semantic information of both time and space dimensions by establishing a relationship network based on correlation, effectively fusing spatiotemporal features to achieve effective modeling of spatiotemporal relationships. Furthermore, this network-based modeling method further supports community discovery based on relationship networks, ultimately providing a method for extracting evolutionary patterns that accurately reflect spatiotemporal relationships.

[0026] 2) Preserving local details and improving the stability of pattern extraction in the time dimension: This method adopts a wrapping window strategy, which considers the correlation between two consecutive time slices in the correlation determination, thereby improving the stability of pattern extraction; while the window wrapping still performs correlation analysis at the granularity of time slices, preserving local details as much as possible.

[0027] 3) Broad spatial pattern detection: Long-duration, wide-ranging spatiotemporal events may link geographically distant time series. Community-based methods for extracting evolutionary patterns eliminate geographical distance limitations on the geographical time series within the patterns, enabling the detection of a wider range of evolutionary patterns. For example, two geographically distant time series may still be considered part of the same evolutionary pattern if they are correlated with most geographical time series within the same pattern. Attached Figure Description

[0028] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0029] Figure 1 This is a flowchart of a method for extracting spatiotemporal evolution patterns from geographic time series, provided in the embodiment.

[0030] Figure 2 This is a schematic diagram of the method for extracting spatiotemporal evolution patterns from geographic time series provided in the embodiment. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and do not limit the scope of protection of this invention.

[0032] In this embodiment, the terminology used is defined as follows: It is a collection of locations, each equipped with sensors (e.g., air quality monitoring stations). Each location has geographic coordinates, and the distance between two locations is denoted as dis(l). a , l b ).

[0033] A time period is a time period with a start timestamp (pt). s and end timestamp pt e The data structure, i.e., p = (t s , t e Both time slices and time windows have this structure.

[0034] It is a series of locations The time series collected in the middle, of which the geographic time series V i ={v i，1 vi，2 ,...v i，n} contains each location l i Recordings are made in chronological order (i.e., sensor readings), with identical time intervals between recordings.

[0035] This embodiment uses V i (p) represents the geographic time series V i In time period p (starting from time period pt) s The time period ends pt e The part of ) that is, V i (p) = {..., v i，t , ...}, where timestamp t∈[pt s pt e ].

[0036] The evolutionary model consists of the following two parts: 1) a set of geographically close geographical time series and 2) a time interval p, such that They have similar temporal characteristics, among which, V i (p) represents V i The subsequence within time period p, Seq sub This is the set of these subsequences. The geographic time series in the evolutionary pattern exhibit spatiotemporal relationships in pairs.

[0037] Figure 1 This is a flowchart illustrating the method for extracting spatiotemporal evolution patterns from geographic time series provided in the embodiment. Figure 1 As shown, the method for extracting spatiotemporal evolution patterns from geographic time series provided by this embodiment of the invention includes the following steps:

[0038] S101: Obtain geographic time series data collected by the sensor, perform time slicing on the geographic time series data to obtain multiple time slices, and execute S102 to S104 for each time slice.

[0039] The time span of the geographic time series is uniformly divided into T time slices of equal size, S = {s1, s2, ... s}. T}, where time slice s i It has a time period structure, s i Let ∈S, where S represents the set of time slices and i represents the time slice index. Geographic time series typically have a long time span, usually measured in years or months. Individual evolutionary patterns, on the other hand, have shorter durations, typically measured in minutes, hours, or days. To extract local features from geographic time series, the continuous time dimension needs to be converted into discrete time slices. The specific time slice size depends on the application scenario; too small a time slice makes it difficult to capture significant features, while too large a time slice may make it difficult to capture local details and dynamics.

[0040] Specifically, the data used in this embodiment comes from China's Air Quality Index (AQI) dataset. The dataset contains AQI values ​​from 448 sensors from January 1, 2018 to June 3, 2018, a total of 184 days. The sensors recorded data every three hours, resulting in a geographic time series containing 4,416 values. The recorded AQI values ​​range from [0, 500]. Air pollution data typically evolves on a daily basis; therefore, one day is chosen as the time slice size. Figure 2 As shown, A represents each s i For a time slice, each time slice s i It contains multiple subsequences of geographic time series (curves).

[0041] For each time slice s i Execute steps S102 to S104:

[0042] S102, based on each time slice, generate a time window using a package window strategy, calculate the correlation between any two geographic time series within each time window, and use a first threshold to filter out multiple geographic time series pairs with strong correlation.

[0043] The Pearson correlation coefficient is the most commonly used measure of correlation. However, subtle sudden events or data measurement errors can lead to instability in the correlation coefficient. For example, when using the Pearson correlation coefficient to measure the correlation between two geographic time series over three consecutive time slices, the two geographic time series may be determined to be correlated in the two consecutive time slices, but uncorrelated in the middle time slice due to measurement errors of individual values. This can result in the correlation of geographic time series changing twice within a very small time span, such as within three time slices, from correlated to uncorrelated and back to correlated. To capture significant and accurate patterns, it is necessary to smooth the temporal correlation of geographic time series. Therefore, this method proposes a wrapping window strategy to avoid such abrupt changes.

[0044] Based on the Pearson correlation coefficient, this method introduces a wrapping window strategy, which includes: first, based on each time slice s... i Generate time window w i .time window w i From the i-th central time slice s i and two adjacent time slices s i-1 s i+1 Composition. Time window w i =(s i-1 .t s s i+1 .te ), where s i-1 .t s This indicates the time slice s preceding the center time slice. i-1 The start timestamp, s i+1 .t e Indicates the next time slice s after the center time slice i+1 The end timestamp. The Pearson correlation coefficient method is used to calculate the w of the i-th time window. i Any two geographic time series V a and V b Correlation coefficient r(V) a V b w i The specific formula is as follows:

[0045]

[0046] Where r represents the Pearson correlation coefficient, and the time window w i Contains three time slices {s i-1 s i s i+1}, V a (w i ) is V a Within the time window w i All values ​​within V b (w i ) is V b Within the time window w i All values ​​within, and V a (w i V b (w i The mean of x i，j V represents a (w i For any value in ), y i，j V represents b (w i Any value from ).

[0047] Introducing the first threshold th r Filter out sets of highly correlated geographic time series pairs Among them, geographic time series pairs (V a V b ) and (V b V a )equivalence.

[0048] Specifically, determine the correlation within the time slice, such as Figure 2 As shown, B represents wi A time window contains three time slices {s} i-1 s i s i+1 For 448 geographic time series within three time slices, with 8 values ​​for each slice and a total of 24 values, the Pearson correlation coefficient was used to determine correlation, with the first threshold being th. r Setting it to 0.7 yields a set of multiple geographic time series pairs, CT. wi The wrapping window strategy used in this embodiment considers the correlation between two consecutive time slices in the correlation determination, thereby improving the stability of pattern extraction. The window wrapping still performs correlation analysis at the granularity of time slices, preserving local details as much as possible.

[0049] S103. Based on geographical time series pairs with strong correlation, a second threshold is used to filter out geographical time series pairs that are spatially close, and a graph relationship network is constructed.

[0050] Within each time slice, a relational network is constructed based on the correlations of geographic time series pairs within the time window. A second threshold, th, is used. d Filter out highly correlated geographic time series pairs (CT) wi The set P of geographically close time series pairs wi ={(V a V b )|dis(l a , l b )<th d , (V a V b )∈CT wi}, where dis(l a , l b )<th d Represents the geographic time series V a and geographic time series V b The distance between them is less than the second threshold th d Based on geographic time series, set P wi Construct a graph relation network, which is represented as an unweighted and undirected graph G. wi ={VX wi P wi}, where VX wi For P wi The set of all geographic time series contained in a geographic time series, i.e., the set of nodes in the graph, P wi The set of edges in the graph is represented in the form of geographic time series pairs. The unweighted, undirected edge between two nodes in the graph represents the geographic time series V. a V b In time slicesi They are internally related and spatially adjacent.

[0051] This step takes into account both geographical proximity and temporal relevance in S102. It integrates temporal and spatial information while preserving their respective semantics.

[0052] Specifically, based on the correlation determination results, i.e., the geographic time series, for P wi Construct an unweighted and undirected graph G wi ={VX wi P wi The set of nodes in the graph is a set of geographic time series, and the set of edges is the set P of geographic time series pairs obtained in S102. wi The graph in this embodiment contains 448 nodes, each representing a geographic time series. Then, based on the geographic distance between geographic time series pairs, a second threshold th is used. d Filtering edges in the network, wherein the second threshold th used in this embodiment d =300km. This can be understood as: the geographical distance between the two endpoints of the edge representing the geographical time series does not exceed the threshold of 300km, that is, the farthest distance that air pollution can travel by wind in a day.

[0053] S104, based on graph relation networks, uses a community detection algorithm to extract evolutionary patterns.

[0054] This embodiment uses the Louvain (Blondel VD, Guillaume JL, Lambotte R, et al. Fast folding of communities in large networks[J]. Journal of statistical mechanics: theory and experiment, 2008, 2008(10): P10008.) algorithm to discover communities (i.e. evolutionary patterns), where each geographic time series can only belong to one community.

[0055] The following is a brief introduction to the principle of the Louvain algorithm. The algorithm consists of two phases and iterates repeatedly. Let G = (V, E) denote the graph input to the Louvain algorithm, where V and E are the set of nodes and edges, respectively. Nodes may represent geographical time series or communities, and (undirected) edges represent the relationships between corresponding nodes. W(V) a V b ) is an edge (V a V b The weights of ) using C. wi (V n Let node V be a delimiter. nThe community (set of nodes) where C is located wi ={C wi (V n )|V n ∈V}.

[0056] Input: On the first run, G = G wi (S103 Chinese Relationship Network), W(V) a V b ) = 1; During the iteration process, G = G c W(V) a V b ) = W c (V a V b ), where G c W c This is the result of the algorithm's previous iteration.

[0057] Phase 1: Traverse each node V n ∈V, calculate when it is from C wi (V n ) The community has removed or added separately to excluding C wi (V n The collection of all other communities except for ) The change in community modularity ΔQ for all communities is given by the formula below (the formula for the former is similar and denoted as ΔQ). remove (V n (This will not be elaborated further):

[0058]

[0059] Bianji

[0060] E in，j ={(V e1 V e2 )|V e1 ∈C j V e2 ∈C j , (V e1 V e2 )∈E}

[0061] E tot，j ={(V e1 V e2 )|V e1 ∈V, V e2 ∈C j , (V e1 V e2 )∈E}

[0062] E n ={(Vn V e )|V e ∈V, (V n V e )∈E}

[0063] E n，j ={(V n V e )|V e ∈C j , (V n V e )∈E}

[0064] Let the weight summation function of the edge set be...

[0065]

[0066] ∑ in the formula in，j =S w (E in，j );∑ tot，j =S w (E tot，j );k n =S w (E n );k n，j =S w (E n，j ); m = S w (E). Let ΔQ add ={ΔQ add (V n C j )|C j ∈Ψ}, if max(ΔQ) add )+ΔQ remove (V n If ) > 0, then let C wi (V n ) = C max C max For ΔQ add The largest community. This phase repeats continuously until a certain traversal is completed.

[0067] Phase 2: Building the new network G c =(C wi E c ),

[0068] Remember W c (C a C b ) = S w ({(V e1 V e2 )|Ve1 ∈C a V e2 ∈C b , (V e1 V e2 )∈E}), E c ={(C a C b )|C a ∈C wi C b ∈C wi W c (C a C b W > 0 c (C a C b The weight of an edge is denoted as .

[0069] Repeat the two phases until the network no longer changes or reaches the set maximum number of iterations.

[0070] Specifically, the Louvain community detection algorithm is used to discover evolutionary patterns, i.e., communities in relational networks. This can be achieved by calling the `sknetwork.clustering.Louvain()` function of `scikit-network` with default parameters, and inputting the network from S103 into the function. For example... Figure 2 As shown, D represents the community (evolutionary pattern) detection process.

[0071] The community-based method for extracting evolutionary patterns used in this embodiment allows the geographic time series within the pattern to be unrestricted by geographic distance, thereby enabling the detection of evolutionary patterns over a wider range. In this embodiment, if two geographic time series that are geographically far apart are related to most geographic time series within the same evolutionary pattern, they can still be considered to belong to the same evolutionary pattern.

[0072] This embodiment preserves semantic information in both time and space dimensions by establishing a relationship network based on correlation, effectively fusing spatiotemporal features and achieving effective modeling of spatiotemporal relationships. Furthermore, this network-based modeling method further supports community discovery based on relationship networks, ultimately providing a method for extracting evolutionary patterns that accurately reflect spatiotemporal relationships.

[0073] The specific embodiments described above illustrate the technical solution and beneficial effects of the present invention in detail. It should be understood that the above description is only the most preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, additions, and equivalent substitutions made within the scope of the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for extracting spatiotemporal evolution patterns from geographic time series, characterized in that, Includes the following steps: Step 1: Obtain geographic time series data collected by sensors, perform time slicing on the geographic time series data to obtain multiple time slices, and perform steps 2 to 4 for each time slice; Step 2: Based on each time slice, generate a time window using a package window strategy, calculate the correlation between any two geographic time series within each time window, and use a first threshold to filter out multiple geographic time series pairs with strong correlation; wherein, strong correlation refers to the correlation coefficient value being in the range of [0.5, 1), and the first threshold is set to 0.7; Step 3: Based on the strongly correlated geographic time series pairs, use a second threshold to filter out the spatially close geographic time series pairs and construct a graph relationship network; wherein, the second threshold is 300km; constructing the graph relationship network includes: based on the set of geographic time series pairs P wi Construct a graph relation network, which is represented as an unweighted and undirected graph G. wi ={VX wi ,P wi }, where VX wi For P wi The set of all geographic time series contained in a geographic time series, i.e., the set of nodes in the graph, P wi The graph is represented by a set of edges in the form of geographic time series pairs, where the unweighted, undirected edge between two nodes represents the geographic time series V. a V b In time slices i They are internally related and spatially adjacent; Step 4: Based on the graph relationship network, the evolutionary patterns are extracted using a community detection algorithm. The community detection algorithm uses the Louvain algorithm to extract the community to which each geographic time series belongs, i.e., the evolutionary pattern. Each geographic time series can only belong to one community.

2. The method for extracting spatiotemporal evolution patterns from geographic time series according to claim 1, characterized in that, In step 1, the process of slicing the geographic time series data into multiple time slices includes: The geographic time series data The time span is uniformly divided into T time slices of equal size s. i s i ∈S, S={s1,s2,…s T }, wherein the time slice s i This is represented by a data structure with a start timestamp and an end timestamp, where S represents the set of time slices and i represents the time slice index.

3. The method for extracting spatiotemporal evolution patterns from geographic time series according to claim 1, characterized in that, In step 2, generating a time window using the wrapping window strategy includes: for the i-th time slice s i Generate time window w i At that time, time slice s i As the central time slice, based on the central time slice s i and two adjacent time slices s i-1 ,s i+1 Generate time window w i =(s i-1 .t s ,s i+1 .t e ), where s i-1 .t s This indicates the time slice s preceding the center time slice. i-1 The start timestamp, s i+1 .t e Indicates the next time slice s after the center time slice i+1 The end timestamp.

4. The method for extracting spatiotemporal evolution patterns from geographic time series according to claim 1, characterized in that, In step 2, calculating the correlation between any two geographic time series within each time window includes: calculating the correlation of the i-th time window w. i Any two geographic time series V a and V b Pearson correlation coefficient r(V) a V b ,w i ).

5. The method for extracting spatiotemporal evolution patterns from geographic time series according to claim 1, characterized in that, In step 2, the first threshold is used to filter out multiple geographical time series pairs (V) with strong correlation. a V b ), Includes: introducing a first threshold th r Filter out sets of highly correlated geographic time series pairs r(V a V b ,w i ) represents the i-th time window w i Inland geographic time series pairs (V a V b The correlation coefficient of ) This represents geographic time series data.

6. The method for extracting spatiotemporal evolution patterns from geographic time series according to claim 5, characterized in that, In step 3, the step of using a second threshold to filter out the geographical time series pairs that are spatially close includes: using the second threshold th d Filter out the set of geographic time series pairs (CT) with strong correlations. wi The set P of geographically close time series pairs wi ={(V a V b )|dis(l a ,l b ) <th d ,(V a V b )∈CT wi }, where dis(l a ,l b ) <th d Represents the geographic time series V a and geographic time series V b The distance between them is less than the second threshold th d .

Images