IAES International Journal of Artificial Intelligence (IJ-AI)

Vol. 14, No.1, February 2025, pp. 397~407

ISSN: 2252-8938, DOI: 10.11591/ijai.v14.i1.pp397-407  397

Dynamic spatio-temporal pattern discovery: a novel grid and

density-based clustering algorithm

Swati Meshram1, Kishor P. Wagh2

1
Department of Computer Science and Engineering, Government College of Engineering, Amravati, India
2
Department of Information Technology, Government College of Engineering, Amravati, India

Article Info ABSTRACT

Article history: Clustering is a robust machine- learning technique for exploration of patterns
based on similarity of elements over multidimensional data. Spatio-temporal
Received Mar 22, 2024 clustering aims to identify target objects to mine spatial and temporal
Revised Jul 9, 2024 dimensions for patterns, regularity, and trends. It has been applied in human-
Accepted Jul 26, 2024 centric applications, such as recommendation systems, urban development
and planning, clustering of criminal activities, traffic planning, and
epidemiology to identify the extent of disease spread. Although the existing
Keywords: research work in the field of clustering relies widely on partition and density-
based methods, no major work has been carried out to handle the
Centroids spatiotemporal dimension and understand the dynamics of temporal variation
Density and connectivity between clusters. To address this, our paper proposes an
Distance algorithm to mine clustering patterns in spatiotemporal dataset using an
Earthquake dataset adaptive, dynamic hybrid technique based on grid and density clustering. We
Neighborhood adopt spatio-temporal partitioning of the virtual grid for distribution of data
and reducing distance computation and increasing efficiency. Grouping the
higher density regions along with neighborhood cluster density attraction rate
to merge the clusters. This method has been experimentally evaluated over the
Indian earthquake dataset and found to be effective with clustering silhouette
index up to 0.93.
This is an open access article under the CC BY-SA license.

Corresponding Author:
Swati Meshram
Department of Computer Science and Eng, Government College of Engineering
Amravati, Maharashtra, India
Email: [email protected]

1. INTRODUCTION
Advancement in computer technology, remote sensing, and location-based services has resulted in the
generation of massive spatiotemporal data. Spatiotemporal data analysis is an emerging research area driven by
development and application of intelligent computational techniques. Analyzing spatiotemporal data is beneficial
for various human-centered applications like recommendation systems, identifying disease outbreak patterns,
urban development clustering, infrastructure planning, and detection of criminal activities. Clustering is a valuable
analysis tool for exploring and understanding rich information contained in the spatio-temporal datasets.
The aim of analyzing spatiotemporal data clustering is to detect and examine noteworthy patterns in the
data that change both in space and time and help in understanding the dynamics or processes driving the patterns
and trends. Anomalous patterns are detected to indicate rare but significant events with deviation from expected
behavior. The analysis may also be used to design models for predicting future occurrences of similar events. It
provides insights in evolution of cluster and its changes over time which is a valuable information in
understanding trends. This may lead to proactive decision-making and improved resource allocation.
Understanding the impact of human activities and natural processes and their interconnections. This is highly

Journal homepage: http://ijai.iaescore.com

398  ISSN:2252-8938

relevant in the field of emergency response, public health, monitoring environmental variables, and urban resource
allocation. Spatiotemporal clustering analysis allows researchers to uncover complex relationships and make
informed decisions in dynamic and interconnected systems [1]. Spatio-temporal data contain the geographical
location and temporal or time of occurrence of the events along with other non-spatiotemporal features describing
the events. Spatio-temporal data clustering analysis is a machine learning technique to search patterns in a dataset
by grouping data instances based on similarity measures. The intraclusters instances exhibit high similarity
whereas the same instances are incoherent with instances of other clusters to form distinct clusters. In other words,
clustering, is an unsupervised classification technique that separates an unlabeled data set into a finite number of
groups whose members are data instances that are more homogeneous to its group than to other groups. These
groups are termed as clusters. Thus, clustering as a process is illustrated by the following example:
Given an input data set X={x1, x2,…,xn}, where each xi has set of j features or dimensions. We attempt
to derive ‘k’ clusters given as C={c1, c2,…ck} satisfying the following conditions. For all i,j ϵ {1,…k},each
|ci| >0, and 𝐶𝑖 ∩ 𝐶𝑗 = ø, and 𝑋 =∪𝑘𝑖=0 𝐶𝑖 .
Clustering has been applied in recommendation systems by reviewing customer feedbacks for product
popularity. Clustering as a tool is also useful for observing the abnormal behavior of outliers that do not exhibit
the same relationship as that of other clusters. This analysis helps detect rare but important patterns in urban
planning [2], big climate data analytics [3]. One of the applications of clustering we tend to explore is detection
of earthquake clusters of same severity, regions of clustering displaying foreshocks and aftershocks of main
earthquake events. Earthquakes are natural events that cause tremors from Earth’s core to the surface. These
sudden, vibrations may destroy useful natural and man-made resources. Identification of such areas, which
may have a trend or reach of earthquake impacts using machine learning pattern mining techniques is
important. Hence, we focus our study on deriving clustering patterns through our proposed research work on
Indian earthquake spatiotemporal data. We highlight our contribution in this research article as follows:
i) a method for selection of centroids; ii) a method to convert tentative clusters to fixed clusters based on
density; iii) outlier score and clustering quality; iv) detecting spatio-temporal referenced variables with respect
to evolution over time; and v) the proposed algorithm is implemented and experimentally validated.
Our research paper adheres to the following structure: section 2 explores related literature. Section 3
outlines the methodology. Section 4 presents the results and subsequent discussion. Finally, section 5 provides
the conclusion of the research.

2. RELATED WORK
The clustering distance-based method computes a distance metric to measure the spatial distance and
cluster similar or neighbouring points. The distance metric used are Euclidean distance, dynamic time warping
[4], longest common subsequence (LCSS) [5], edit distance on real sequence (EDR) [6], Hausdroff [7], and
Fréchet [8] distance. Density based clustering performs the grouping of density satisfying regions into clusters
[9]. Feature-based clustering, first extracts the features and then computes their similarity [10], [11]. Time
series data analysis using kernel density was the work undertaken to develop the algorithm spatio-temporal
density-based spatial clustering of applications with noise (ST-DBSCAN) [12]. A spike neural network
architecture is developed to cluster spatiotemporal brain data [13]. Guo et al. [14] analysed foodborne diseases
on people of Zhejiang has been studied using spatiotemporal clustering which includes methods such as
statistical and spatial analysis along with spatiotemporal scanning. Here the temporal resolution found is large.
Loiola et al. [15] explored a hybrid burned area algorithm based on moderate resolution imaging
spectroradiometer (MODIS) thermal anomalies and NIR reflectance with spatial resolution of 250 m on
MODIS data. Hotspots clusters were developed to discover fire active areas. All these studies reflect the
algorithms are developed to tackle specific problems and thus they have limited applicability. Gong et al. [16]
put forth a model that learns the dynamics of mobility in taxi trajectory data and uses it to predict mobility in
specific route areas. Here the area of study is confined to a particular area. Another article on trajectory
clustering is studied in [17], it identifies clusters based on Hausdroff distance in K-nearest neighbour method
where the accuracy of the method heavily relies on appropriate value of ‘K’. The research in [18], [19]
trajectory analysis was used to extract road traffic, determine flow statistic, and detect congestion. According
to Georgoulas et al. [20], a hybrid approach of clustering over seismic spatio-temporal data was adopted. It is
based on density and hierarchical agglomerative clustering which extracts objects with unknown class labels.
Here connectivity is based on single linkage to form the clusters and no emphasis is placed on the temporal
parameter of the dataset. According to Nazia et al. [21], space time clusters were discovered using
geographically weighted regression model. The model also utilizes local and global Moran’s I to interpret the
cluster distribution pattern. The model was verified using COVID dataset. Another work on COVID dataset is
carried out in [22]. The authors adopted a partition dataset using medoids and improved the result gap statistics.
While some of these studies have discussed about outliers, but they have not explicitly addressed to reduce the

Int J Artif Intell, Vol. 14, No.1, February 2025: 397-407

Int J Artif Intell ISSN:2252-8938  399

outlier’sratio. A comparison of different types of spatiotemporal clustering is presented in Table 1. It highlights

recent works in spatio-temporal clustering analysis with its applications and limitations of methods. However,
these conventional approaches have limitations. It is a less studied topic and the impact of temporal resolution
influencing the spatial events has not been thoroughly studied.

Table 1. Recent work on spatio-temporal clustering technique

Reference Method category Method/model name Model validation Application Limitations in the
article/method
[21] Spatial regression Geographical AIC, R2, Log COVID-19 cluster Outliers can have
and space-time weighted regression likelihood analysis disproportionate impact on
scan statistics model prediction.
[22] spatial auto K-medoid, Spectral Gap Statistics COVID-19 cluster Computational complexity is
correlation density matrix analysis higher.
[23] agglomerative - accuracy - 92-96%, Drought Analysis Uses NLP bag of words to
hierarchical recall and precision capture the location which is
clustering imprecise.
[24] agglomerative Ward system minimum variance Distribution of social Single factor detection
hierarchical clustering enterprises across analysis, multiple features
clustering provinces in China not included.
[25] spatio-temporal Kulldorff’s space-time log likelihood ratio test Hotspot detection of Factors triggering cluster
Clustering scan statistic, discrete COVID-19 cases in formation were unclear.
Poisson model Johor, Malaysia.
[26] Deep learning MuSTC, spatial MAE - 0.2304, RMSE- Sea Surface Temperature Additionally, Regional
correlation 0.3527 prediction information required.
[27] Statistical The Gertis Ord Gi* for p-value<0.05, z-score disease control-Bovine Experiments conducted on
hotspot analysis >1.96 with confidence anaplasmosis across data of two years. Shorter
level 90,95,99% Zimbabwe period.
[28] Hierarchical Average Linkage Gap Statistics and Drought analysis of Study doesnot include full
clustering criterion, partitioning Silhouette methods, p- Lowveld inthe Limpopo range of dependence of
around medoids, value>0.05, Hopkins Province in South Africa. parameters. Relied on partial
Smith and Schlather statistic>0.5, extremal dependence of
model, Hopkins parameters.
statistic
[29] Space-time scan Binomial regression p-value<0.05, monte- Clustered attacks of Cannot detect irregular shape
statistics model-Poisson model. carlo likelihood leopards on humans in clusters
Himachal Pradesh, India
2004-2015
[30] Density based and DBSCAN, multiscale Akaike information spatial agglomeration of spatial relationship between
regression model geographically criterion (AIC) the catering industry the service industry and
weighted regression residential is explored and
(MGWR) not explored other industry.
[31] Hierarchical Self-organized maps Adjusted randomized - The SOM architecture
clustering (SOM) index (ARI), requires modification for
Silhouette score, different dataset.
Calinski–Harabasz Experiments carried on
score standard dataset from UCI
repository.
[32] Space-time scan Kulldorff’s space-time Loglikelihood ratio spatiotemporal clusters Study is confined to a specific
statistics scan statistic, discrete (LLR), Relative Risk of malaria incidences region dataset. Cannot detect
Poisson model (RR), p<0.05, LLR= irregular shape clusters
886,097.7, RR= 5.55,
P< 0.05
[33] Spatio-temporal discrete Poisson LLR, RR, Monte- cluster of human Study is confined to a specific
scan statistics model, window scan Carlo statistical brucellosis region dataset. Cannot detect
significance. irregular shape clusters

3. METHODOLOGY
This section discusses the proposed methodology using hybrid grid and density-based clustering
approach on spatio-temporal data as shown in the Figure 1. The proposed method is implemented using Python
programming in Colab environment which offers Google free cloud space for storage of data along with CPU
processing capability. We have employed the concept of grid structure, density of the grid cells and
neighborhood of instances, centroids and grids to form clusters. Further use density attraction rate of
neighboring clusters to merge the clusters and derive final clusters.
The dataset is obtained from the https://seismo.gov.in [34] which is the Government of India portal
for Seismic events containing earthquake spatiotemporal data for the Indian subcontinent. The data is available
in CSV file. With 6506 samples have been employed in our experiment from the year August 2019 to January
2024 as mentioned in Table 2. The attributes of the dataset are spatial longitude, latitude, timestamp, and depth
of the event along with comments. Table 3, shows the types of earthquake severity levels based on magnitude
Dynamic spatio-temporal pattern discovery: a novel grid and density-based … (Swati Meshram)
400  ISSN:2252-8938

and depth. We imported the earthquake catalogue and deleted the comments describing the textual location of
the earthquake. Table 4, describes the parameters of the proposed algorithm with its initialization. The
visualisation of different levels of earthquakes recorded are shown in Figures 2(a) and 2(b).

Figure 1. The proposed spatio-temporal clustering framework

Table 2. Spatio-temporal dataset and instances

Dataset Total instances
Indian Earthquake Catalogue August 2019 to January 2024 6506

Table 3. Descriptive statistics of earthquake dataset

Attribute Min Max Mean Median
Magnitude 1.2 7 4.4 3.9
Depth 0.8 471 48 10

Table 4. Parameters of the algorithm with initialization

Parameter Description Initializationand range
λ Total equally spaced cells 10,20
𝛾 Minimum probable centroids in a grid cell 2
r Increment rate 1 to λ
minDensity Minimum density of a cluster 10
minTH Minimum Threshold 0.1

(a) (b)

Figures 2. Summary of different w.r.t count of earthquake dataset based on (a) magnitude and
(b) depth levels

Spatiotemporal data is a sequence of data points in increasing order of time and is expressed as (1).

𝑆𝑇 = {𝑠𝑡1 , 𝑠𝑡2 , … . . , 𝑠𝑡𝑛 } (1)

where ST is a collection of spatio-temporal events dataset and ‘n’ is the total number of spatiotemporal events
present in the dataset. sti- represents the ith datapoint that records the longitude, latitude as location coordinates
along with occurrence time of the event.
Grid: G is a multidimensional logical grid that geographically and temporally divides the
spatiotemporal space. The division is based on the longitude, latitude, and time. In addition, the dataset also
records non-spatial information related to events.
Distance Measure: This measure describes the closeness of two data points based on their spatial-
temporal distance and similarity, producing lower values for low similarity and higher values for high

Int J Artif Intell, Vol. 14, No.1, February 2025: 397-407

Int J Artif Intell ISSN:2252-8938  401

similarity. As a spatio-temporal distance measure, we adopt the Haversine distance formula for spatial distance
along with the temporal distance measured in days. We assume that the time advancement between any two
events was at least 1. The Haversine distance formula for two spatial instances is expressed as (2):

𝐻𝑑𝑖𝑠𝑡(𝑂𝑖 , 𝑂𝑗 ) = 2𝑅. 𝑎𝑟𝑐𝑠𝑖𝑛 √𝑠𝑖𝑛2 ( 𝐵 − 𝐴) + 𝑐𝑜𝑠( 𝐴). 𝑐𝑜𝑠( 𝐵) + 𝑠𝑖𝑛2 ( 𝐷 − 𝐶) (2)

where R is radius of the Earth with value as 6371 kms. A, B, C, D represents:

𝐴 = 𝑂𝑖 . 𝑙𝑎𝑡, 𝐵 = 𝑂𝑗 . 𝑙𝑎𝑡, 𝐶 = 𝑂𝑖 . 𝑙𝑜𝑛, 𝐷 = 𝑂𝑗 . 𝑙𝑜𝑛 (3)

𝑡𝑜_𝑑𝑎𝑦𝑠(𝑂𝑖 . 𝑡𝑖𝑚𝑒 − 𝑂𝑗 . 𝑡𝑖𝑚𝑒), 𝑖𝑓(𝑂𝑖 . 𝑡𝑖𝑚𝑒 ≠ 𝑂𝑗 . 𝑡𝑖𝑚𝑒)

𝑇𝑑𝑖𝑠𝑡(𝑂𝑖 , 𝑂𝑗 ) = { (4)
1, 𝑖𝑓(𝑂𝑖 . 𝑡𝑖𝑚𝑒 = 𝑂𝑗 . 𝑡𝑖𝑚𝑒)

Step 1: Determine minimum and maximum longitude and latitude coordinates of the dataset.

𝐺(min𝑙𝑜𝑛, min𝑙𝑎𝑡) = ∀𝑖 min(𝑠𝑡𝑖 . 𝑙𝑜𝑛𝑔𝑖𝑡𝑢𝑑𝑒, 𝑠𝑡𝑖 . 𝑙𝑎𝑡𝑖𝑡𝑢𝑑𝑒)

𝐺(max𝑙𝑜𝑛, max𝑙𝑎𝑡) = ∀𝑖 𝑚𝑎𝑥 ( 𝑠𝑡𝑖 . 𝑙𝑜𝑛𝑔𝑖𝑡𝑢𝑑𝑒, 𝑠𝑡𝑖 . 𝑙𝑎𝑡𝑖𝑡𝑢𝑑𝑒 (5)

𝑡𝑚𝑖𝑛 = ∀𝑖 min(𝑠𝑡𝑖 . 𝑡𝑖𝑚𝑒)

𝑡𝑚𝑎𝑥 = ∀𝑖 max(𝑠𝑡𝑖 . 𝑡𝑖𝑚𝑒) (6)

Step 2: Calculate the step size.

(maxlat-minlat)
𝛥lat=
𝜆
(maxlon-minlon)
𝛥Lon=
𝜆
(𝑡𝑚𝑎𝑥 −𝑡𝑚𝑖𝑛 )
∆𝑡 = |𝑆𝑇|
(7)

where λ is initialised to 20.

Step 3: Form spatio-temporal grid with grid cells.
𝐿𝑎𝑡1 = minlat + 𝑟 × 𝛥𝐿𝑎𝑡, 𝑟 ≤ 𝜆
𝐿𝑎𝑡2 = 𝐿𝑎𝑡1 + 𝛥𝐿𝑎𝑡, 𝐿𝑎𝑡2 ≤ 𝑚𝑎𝑥𝑙𝑎𝑡
𝐿𝑜𝑛1 = 𝑚𝑖𝑛𝑙𝑜𝑛 + 𝑟 × 𝛥𝐿𝑜𝑛, 𝑟 ≤ 𝜆
𝐿𝑜𝑛2 = 𝐿𝑜𝑛1 + 𝛥𝐿𝑜𝑛, 𝐿𝑜𝑛2 ≤ 𝑚𝑎𝑥𝑙𝑜𝑛 (8)

𝐺𝑠𝑝𝑎𝑡𝑖𝑎𝑙(𝐿𝑎𝑡1 , 𝐿𝑎𝑡2 , 𝐿𝑜𝑛1 , 𝐿𝑜𝑛2 ); 𝐺𝑡𝑒𝑚𝑝𝑜𝑟𝑎𝑙(𝑡𝑗, 𝑡𝑗 + 𝛥𝑡) (9)

Step 4: Allocate the data-points to the grid cells.

𝐺𝑖𝑗𝑘 = {𝑜𝑚 |(𝐺𝑖𝑗𝑘 . minlat ≤ 𝑜𝑚 . 𝑙𝑎𝑡 ≤ 𝐺𝑖𝑗𝑘 . maxlat) ∧ (𝐺_𝑖𝑗𝑘. 𝑚𝑖𝑛𝐿𝑜𝑛. ≤ 𝑜_𝑚. 𝑙𝑜𝑛 ≤ 𝐺_𝑖𝑗𝑘. 𝑚𝑎𝑥𝑙𝑜𝑛)
∧ 𝑜𝑚 . 𝑡𝑖𝑚𝑒 ∈ [𝑘 − 1, 𝑘]} (10)

where Om represents the spatio-temporal instances, with its event time belonging to the interval k-1 to k.
Step 5: Compute the density of each grid cell as

𝐷𝑒𝑛𝑠𝑖𝑡𝑦(𝐺𝑖𝑗𝑘 ) = |𝑂𝑚 |, ∀𝑂𝑚 ∈ 𝐺𝑖𝑗 ∧ 𝑂𝑚 . 𝑡𝑖𝑚𝑒 ∈ [𝑘 − 1, 𝑘] (11)

where Gijk represents the grid cell index number.

Step 6: Determine ‘p’, the maximum number of probable centroids.

𝑝 = |𝐿𝑜𝑔2 (𝐷𝑒𝑛𝑠𝑖𝑡𝑦(𝐺𝑖𝑗𝑘 )) + 𝛾| (12)

where γ takes value as 2.

Step 7: Distribution of probable centroids (PCentre).

If 𝐷𝑒𝑛𝑠𝑖𝑡𝑦(𝐺𝑖𝑗𝑘 ) > 𝑚𝑖𝑛𝐷𝑒𝑛𝑠𝑖𝑡𝑦

/* Select probable centroids P */
𝑃𝐶𝑒𝑛𝑡𝑟𝑒(𝐺𝑖𝑗𝑘 ) = {𝑅𝑎𝑛𝑑𝑜𝑚(𝑂𝑚 )|𝑂𝑚 ∈ 𝐺𝑖𝑗𝑘 } (13)

|𝑃𝐶𝑒𝑛𝑡𝑟𝑒(𝐺𝑖𝑗𝑘 )| ≤ 𝑝 (14)

Dynamic spatio-temporal pattern discovery: a novel grid and density-based … (Swati Meshram)
402  ISSN:2252-8938

where p is the maximum number of probable centroids.

Step 8: Assign each grid cell data instances to the closest probable centroids with minimum distance to form
probable clusters.

𝑃𝐶𝑙𝑢𝑠𝑡𝑒𝑟𝑞 (𝑂𝑚 ) = 𝑎𝑟𝑔𝑚𝑖𝑛𝑞=1..𝑝 {𝑑𝑖𝑠𝑡(𝑂𝑚 , 𝑃𝐶𝑒𝑛𝑡𝑟𝑒𝑞 )} (15)

where dist is the spatio-temporal distance given as

𝑑𝑖𝑠𝑡(𝑂𝑚 , 𝑃𝐶𝑒𝑛𝑡𝑟𝑒𝑞 ) = 𝐻𝑑𝑖𝑠𝑡(𝑂𝑚 , 𝑃𝐶𝑒𝑛𝑡𝑟𝑒𝑞 ) ∗ 𝑇𝑑𝑖𝑠𝑡(𝑂𝑚 , 𝑃𝐶𝑒𝑛𝑡𝑟𝑒𝑞 ) (16)

Hdist is the Haversine spatial distance between two locations. Tdist is the temporal distance between the two
events converted into days.
Step 9: Compute the average radius of the probable clusters.
∑𝑂𝑚∈𝑃𝐶𝑙𝑢𝑠𝑡𝑒𝑟 𝑑𝑖𝑠𝑡(𝑂𝑚 ,𝑃𝐶𝑒𝑛𝑡𝑟𝑒𝑞 )
𝑞
𝑅𝑎𝑑𝑖𝑢𝑠(𝑃𝐶𝑙𝑢𝑠𝑡𝑒𝑟𝑞 ) = (17)
|𝑃𝐶𝑙𝑢𝑠𝑡𝑒𝑟𝑞 |

Step 10: Calculate the density attraction rate of probable centroids.

|𝑃𝐶𝑙𝑢𝑠𝑡𝑒𝑟𝑞 |
𝐷𝑒𝑛𝑠𝑖𝑡𝑦𝐴𝑡𝑡𝑟𝑎𝑐𝑡𝑜𝑛𝑅𝑎𝑡𝑒(𝑃𝐶𝑒𝑛𝑡𝑟𝑒𝑞 ) = (18)
|𝐺𝑖𝑗𝑘 |

Step 11: Sort the density attraction rate of each grid cell.
The density attraction rate is used to merge the clusters with minimum points joins to strong clusters in the
neighbourhood.

Step 12:𝐹𝑜𝑟 𝑒𝑎𝑐ℎ 𝑔𝑟𝑖𝑑 𝑐𝑒𝑙𝑙 𝐺𝑖𝑗𝑘 ∈ 𝐺 𝑑𝑜

𝐹𝑜𝑟 𝑒𝑎𝑐ℎ 𝑃𝐶𝑒𝑛𝑡𝑟𝑒𝑞 ∈ 𝐺𝑖𝑗𝑘 𝑑𝑜
𝐼𝑓(𝐷𝑒𝑛𝑠𝑖𝑡𝑦𝐴𝑡𝑡𝑟𝑎𝑐𝑡𝑖𝑜𝑛𝑅𝑎𝑡𝑒(𝑃𝐶𝑒𝑛𝑡𝑟𝑒𝑞 ) ≤ 𝑚𝑖𝑛 𝑇𝐻 ) 𝑡ℎ𝑒𝑛
𝑃𝐶𝑒𝑛𝑡𝑟𝑒 ← 𝑃𝐶𝑒𝑛𝑡𝑟𝑒 − 𝑃𝐶𝑒𝑛𝑡𝑟𝑒𝑞 (19)

𝐼𝑓(𝐷𝑒𝑛𝑠𝑖𝑡𝑦𝐴𝑡𝑡𝑟𝑎𝑐𝑡𝑖𝑜𝑛𝑅𝑎𝑡𝑒(𝑃𝐶𝑒𝑛𝑡𝑟𝑒𝑞 ) ≥ 0.3) 𝑡ℎ𝑒𝑛

/* Include the cluster in final list of clusters */
𝐶𝑙𝑢𝑠𝑡𝑒𝑟 ← 𝑃𝐶𝑙𝑢𝑠𝑡𝑒𝑟𝑞 (20)
𝐼𝑓(𝐷𝑒𝑛𝑠𝑖𝑡𝑦𝐴𝑡𝑡𝑟𝑎𝑐𝑡𝑖𝑜𝑛𝑅𝑎𝑡𝑒(𝑃𝐶𝑒𝑛𝑡𝑟𝑒𝑞 ) < 0.3) 𝑡ℎ𝑒𝑛
Go to step 13

Step 13: Construct centroid to centroid distance matrix, p x p for p centroid configuration.

𝐸𝑥𝑡𝑒𝑟𝑖𝑜𝑟𝑑𝑖𝑠𝑡(𝑃𝐶𝑒𝑛𝑡𝑟𝑒𝑞 , 𝑃𝐶𝑒𝑛𝑡𝑟𝑒𝑟 ) = 𝑑𝑖𝑠𝑡(𝑃𝐶𝑒𝑛𝑡𝑟𝑒𝑞 , 𝑃𝐶𝑒𝑛𝑡𝑟𝑒𝑟 )−𝑅𝑎𝑑𝑖𝑢𝑠(𝑃𝐶𝑒𝑛𝑡𝑟𝑒𝑞 ) − 𝑅𝑎𝑑𝑖𝑢𝑠(𝑃𝐶𝑒𝑛𝑡𝑟𝑒𝑟 ) (21)

Step 14: Find the neighbour clusters using exterior distance.

𝑁𝑒𝑖𝑔ℎ𝑏𝑜𝑢𝑟(𝑃𝐶𝑒𝑛𝑡𝑟𝑒𝑞 ) = 𝑚𝑖𝑛 𝐸 𝑥𝑡𝑒𝑟𝑖𝑜𝑟𝑑𝑖𝑠𝑡(𝑃𝐶𝑒𝑛𝑡𝑟𝑒𝑞 − 𝑃𝐶𝑒𝑛𝑡𝑟𝑒𝑟 ) (22)

Step 15: Find the neighbour cluster density attraction rate to merge the cluster.

𝐼𝑓 (𝐷𝑒𝑛𝑠𝑖𝑡𝑦𝐴𝑡𝑡𝑟𝑎𝑐𝑡𝑖𝑜𝑛𝑅𝑎𝑡𝑒(𝑃𝐶𝑒𝑛𝑡𝑟𝑒𝑞 ) ≤ 𝐷𝑒𝑛𝑠𝑖𝑡𝑦𝐴𝑡𝑡𝑟𝑎𝑐𝑡𝑖𝑜𝑛𝑅𝑎𝑡𝑒(𝑃𝐶𝑒𝑛𝑡𝑟𝑒𝑟 )) 𝑡ℎ𝑒𝑛

𝑃𝐶𝑙𝑢𝑠𝑡𝑒𝑟𝑟 ← 𝑃𝐶𝑙𝑢𝑠𝑡𝑒𝑟𝑟 ∪ 𝑃𝐶𝑙𝑢𝑠𝑡𝑒𝑟𝑞
𝑃𝐶𝑙𝑢𝑠𝑡𝑒𝑟 ← 𝑃𝐶𝑙𝑢𝑠𝑡𝑒𝑟 − 𝑃𝐶𝑙𝑢𝑠𝑡𝑒𝑟𝑞
𝑃𝐶𝑒𝑛𝑡𝑟𝑒 ← 𝑃𝐶𝑒𝑛𝑡𝑟𝑒 − 𝑃𝐶𝑒𝑛𝑡𝑟𝑒𝑞 (23)

Continue step 12
𝐼𝑓(𝐷𝑒𝑛𝑠𝑖𝑡𝑦𝐴𝑡𝑡𝑟𝑎𝑐𝑡𝑖𝑜𝑛𝑅𝑎𝑡𝑒(𝑃𝐶𝑒𝑛𝑡𝑟𝑒𝑞 ) > 𝐷𝑒𝑛𝑠𝑖𝑡𝑦𝐴𝑡𝑡𝑟𝑎𝑐𝑡𝑖𝑜𝑛𝑅𝑎𝑡𝑒(𝑃𝐶𝑒𝑛𝑡𝑟𝑒𝑟 ))𝑡ℎ𝑒𝑛
𝑃𝐶𝑙𝑢𝑠𝑡𝑒𝑟𝑞 ← 𝑃𝐶𝑙𝑢𝑠𝑡𝑒𝑟𝑞 ∪ 𝑃𝐶𝑙𝑢𝑠𝑡𝑒𝑟𝑟
𝑃𝐶𝑙𝑢𝑠𝑡𝑒𝑟 ← 𝑃𝐶𝑙𝑢𝑠𝑡𝑒𝑟 − 𝑃𝐶𝑙𝑢𝑠𝑡𝑒𝑟𝑟
𝑃𝐶𝑒𝑛𝑡𝑟𝑒 ← 𝑃𝐶𝑒𝑛𝑡𝑟𝑒 − 𝑃𝐶𝑒𝑛𝑡𝑟𝑒𝑟 (24)
Continue step 12

Step 15: Find outlier ratio.

|∪∀𝑞 𝐶𝑙𝑢𝑠𝑡𝑒𝑟𝑞|
𝑂𝑢𝑡𝑙𝑖𝑒𝑟_𝑟𝑎𝑡𝑖𝑜 = |𝑆𝑇|
(25)

Int J Artif Intell, Vol. 14, No.1, February 2025: 397-407

Int J Artif Intell ISSN:2252-8938  403

Step 16: Calculate silhouette index for quality.

𝜇−𝑀
𝑆𝑖𝑙ℎ𝑜𝑢𝑡𝑡𝑒_𝑖𝑛𝑑𝑒𝑥 = (26)
𝑚𝑎𝑥(𝜇,𝑀)

where 𝜇is the mean distance from the centroid to all other data instances within the cluster. M is the mean
distance to all other clusters data instances.
Step 17: Stop.

4. RESULTS AND DISCUSSION

Figures 3 and 4 present the experimental result of proposed hybrid clustering algorithm. The results
display the magnitude and depth of every cluster with respect to time in the Indian subcontinent. This reflects
the density of the formed clusters in space and time dimensions. Densely populated cluster 25 is in Fayzabad,
Afghanistan, Pakistan, and Jammu Kashmir regions of India with 1219 events and a mean magnitude of
4.15 richter scale. The size of cluster 25 has been the highest as reflected in Figure 4. It is observed that the
recurrence duration given by the mean time between the events is 99 hours. The next highly populated cluster
7, is in the eastern India region that include Mizoram, Arunachal Pradesh, and Manipur. The mean magnitude
is 3.48 richter scale. Then the next highest being the Himachal Pradesh, Uttarakhand, and certain region of
Jammu Kashmir as cluster 19 with mean magnitude of 3.08 richter scale. Next is the Andaman Nicobar Islands
cluster 1 with a mean magnitude of 4.45 richter scale. The data analysis also reflects that there are fewer
earthquakes with higher magnitudes and many number of it with moderate magnitudes i.e. majorly the events
fall in the range from 3.5 to 5. It is observed that the stronger earthquakes are followed by lesser magnitude
earthquake in the near surrounding regions. The aftershocks are even felt after several days from the stronger
earthquakes. The outlier ratio was 0.005% towards the 38 events not closer to any of the clusters in space-time
dimension which is constrained by the distance threshold. Outliers are events that belong to regions such as
Oman and Maldives, that have not been in the range of spatio-temporal grid and could not be assigned to any
clusters, due to the spatial distance being greater than the distance threshold value.
The analysis has shown spatial and temporal interactions, and changing the resolution, provides an
effective algorithm for earthquake modeling. If the distance and time thresholds are maximum, will result in
merging many clusters in one or could result in overlap of clusters. The study confirms that magnitude and
frequency are correlated in spatiotemporal dimension and tends to generate clusters. Across and within the
cluster distance and variation differentiate regions with high-risk earthquake zones. Low-risk clusters appear
in the region. The Andaman, Nicobar, Jammu and Kashmir regions comprise high-risk clusters. High-risk
events with magnitudes above six richter scales were observed in cluster 24. Greater depth events are identified
in cluster 25 that make it high-risk clusters. These are highly complex regions due to many fault lines
periodically releasing the tectonic stresses in the form of earthquakes.
The model produces better result for earthquake classification model. The results of the hybrid spatio-
temporal clustering are best due to the accuracy of the results is higher. The advantage that is observed is the
allocation of data instances to grids reduces the burden of comparison to farthest centroids not in their
neighborhood. Eliminating the unnecessary computations. Given that centroid selection is a random process, may
result into increase in the mean intracluster distance. For even distribution of centroids to form clusters with strong
connectivity, the distribution of probable centroids, computation of density attraction and merging of clusters
locally and globally have led to better clustering quality. We used the discard policy to eliminate the clusters and
merge to those neighbor clusters with higher density attraction rate has shown the proposed method is adaptive.
The proposed method in this study tended to have an inordinately higher proportion of data clusters
based spatial and temporal density, which is in contrast to the result as shown in STK-means methods as shown
in the Figures 5 and 6. STK-means clusters are time slice of data instances irrespective to spatial distance
whereas Figure 6 shows non overlapping distinct clusters.Our study suggests that higher number of clusters
formed due to the spatial dimensions of the dataset and density attraction ratio, results into reducing the outliers
and increase in the clustering quality. Which is evident from the result of ST_DBSCAN and proposed method
shown in the Figures 7 and 8. The proposed methods benefits from forming clusters based on centroid selection
for the grid cell depending on its density. The results also demonstrates that the minthreshold parameter minTH
is crucial in determining the outlier threshold. If the minthreshold is close to zero, will result in assignment of
every data instance to some cluster. But even a slight increase in minthreshold, start to increase the outlier
proportion. Further the other parameter that also is of importance is the connectivity linkage distance between
clusters. As this distance between clusters increases, it results into more tightly coupled clusters with higher
density towards the core. On the contrary when the connectivity linkage between the clusters is made small,
the clusters are more arbitrary in shape, and are loosely coupled to the core centroids as the distance between
centroid and the data instances on the border’s increases. The silhouette index is about 0.93 reflecting good

Dynamic spatio-temporal pattern discovery: a novel grid and density-based … (Swati Meshram)
404  ISSN:2252-8938

clustering result. It is observed that on this dataset the proposed algorithm has shown better clustering quality.
With Davis Bouldin DB index 2.337 as shown in the Table 5. The number of clusters as are increased, the
spatio-temporal distances between the clusters are reduced, forming strong clusters.

Figure 3. Result of proposed clustering algorithm on the earthquake dataset showing distinct density of the
clusters with magnitude and depth with respect to time plot

Int J Artif Intell, Vol. 14, No.1, February 2025: 397-407

Int J Artif Intell ISSN:2252-8938  405

Figure 4. Result of proposed clustering algorithm on the earthquake dataset showing trend of events with
clustering size

Figure 5. Result of STK-means on Indian subcontine Figure 6. Result of proposed clustering algorithm on
ntearthquake dataset producing seven clusters Indian earthquake dataset producing seven clusters

Figure 7. Result of STDBSCAN on Indian Figure 8. Proposed algorithm on Indian earthquake

earthquake dataset with outliers shown in black dataset, min_distance_threshold=500 km,
color. Epsilon1=1.15, Epsilon2=500 km, min_density=10, forming 24 clusters and outliers
min_samples=10, forming 19 clusters shown in violet color

Dynamic spatio-temporal pattern discovery: a novel grid and density-based … (Swati Meshram)
406  ISSN:2252-8938

Table 5. Clustering quality index of proposed method

Evaluation Parameter Result
Silhoutte Index 0.93
Davis Bouldin Index 2.337

5. CONCLUSION
This paper proposes a novel and adaptive method of clustering. The method has been experimentally
evaluated on real and standard earthquake dataset of Indian subcontinent. The clustering technique uses grid
and density-based partitioning of data instances. Restricting the analysis to the effects of space and time,
provides us with the information that events that are high intensity events are followed by weak events in the
same clustering region reasoning to aftershocks. Our proposed method has found distinct, non-overlapping
arbitrary shaped clusters on spatial and temporal data with reducing outlier ratio and distance metric
computation by taking advantage of grid structure. The silhouette index is about 0.93 shows good clustering
result. The proposed method for Spatio-temporal clustering is experimented on earthquake dataset but it can
be applied on other Spatio-temporal dataset to study the dynamics of data. Further research direction we would
take up is to minimize the parameter required for the method.

REFERENCES
[1] S. Meshram and K. P. Wagh, “Mining intelligent spatial clustering patterns: A comparative analysis of different approaches,” in
Proceedings of the 2021 8th International Conference on Computing for Sustainable Global Development, INDIACom 2021, 2021,
pp. 325–330, doi: 10.1109/INDIACom51348.2021.00056.
[2] Y. Zheng, L. Capra, O. Wolfson, and H. Yang, “Urban Computing,” ACM Transactions on Intelligent Systems and Technology,
vol. 5, no. 3, pp. 1–55, Oct. 2014, doi: 10.1145/2629592.
[3] F. Hu et al., “ClimateSpark: An in-memory distributed computing framework for big climate data analytics,” Computers and
Geosciences, vol. 115, pp. 154–166, 2018, doi: 10.1016/j.cageo.2018.03.011.
[4] D. Berndt and J. Clifford, “Using dynamic time warping to find patterns in time series,” in Proceedings of the 3rd International
Conference on Knowledge Discovery and Data Mining, 1994, vol. 398, pp. 359–370.
[5] M. Vlachos, G. Kollios, and D. Gunopulos, “Discovering similar multidimensional trajectories,” in Proceedings 18th International
Conference on Data Engineering, pp. 673–684, doi: 10.1109/ICDE.2002.994784.
[6] L. Chen, M. T. Özsu, and V. Oria, “Robust and fast similarity search for moving object trajectories,” in Proceedings of the 2005
ACM SIGMOD International Conference on Management of Data, Jun. 2005, pp. 491–502, doi: 10.1145/1066157.1066213.
[7] B. Guan, L. Liu, and J. Chen, “Using relative distance and hausdorff distance to mine trajectory clusters,” TELKOMNIKA
Indonesian Journal of Electrical Engineering, vol. 11, no. 1, 2013, doi: 10.11591/telkomnika.v11i1.1877.
[8] M. M. Fréchet, “Sur quelques points du calcul fonctionnel,” Rendiconti del Circolo Matematico di Palermo, vol. 22, no. 1, pp. 1–
72, 1906, doi: 10.1007/BF03018603.
[9] Z. Cheng, L. Jiang, D. Liu, and Z. Zheng, “Density based spatio-temporal trajectory clustering algorithm,” International Geoscience
and Remote Sensing Symposium (IGARSS), vol. 2018, pp. 3358–3361, 2018, doi: 10.1109/IGARSS.2018.8517434.
[10] J. Du and L. Aultman-Hall, “Increasing the accuracy of trip rate information from passive multi-day GPS travel datasets: Automatic
trip end identification issues,” Transportation Research Part A: Policy and Practice, vol. 41, no. 3, pp. 220–232, 2007, doi:
10.1016/j.tra.2006.05.001.
[11] N. Pelekis, I. Kopanakis, I. Ntoutsi, G. Marketos, and Y. Theodoridis, “Mining trajectory databases via a suite of distance operators,”
in 2007 IEEE 23rd International Conference on Data Engineering Workshop, Apr. 2007, pp. 575–584, doi:
10.1109/ICDEW.2007.4401043.
[12] D. Birant and A. Kut, “ST-DBSCAN: An algorithm for clustering spatial-temporal data,” Data and Knowledge Engineering, vol.
60, no. 1, pp. 208–221, 2007, doi: 10.1016/j.datak.2006.01.013.
[13] M. G. Doborjeh and N. Kasabov, “Dynamic 3D clustering of spatio-temporal brain data in the NeuCube spiking neural network
architecture on a case study of fMRI data,” in Neural Information Processing, vol. 9492, 2015, pp. 191–198, doi: 10.1007/978-3-
319-26561-2_23.
[14] J.-X. Guo, T. Liu, X.-J. Qi, J. Chen, and S.-Y. Ai, “Application of spatio-temporal scanning in the analysis of spatio-temporal
clusters of foodborne diseases in Zhejiang Province,” Chinese Preventive Medicine, vol. 21, no. 11, pp. 1171–1177, 2020, doi:
10.16506/j.1009-6639.2020.11.003.
[15] J. L. -Loiola, G. Otón, R. Ramo, and E. Chuvieco, “A spatio-temporal active-fire clustering approach for global burned area mapping
at 250 m from MODIS data,” Remote Sensing of Environment, vol. 236, Jan. 2020, doi: 10.1016/j.rse.2019.111493.
[16] S. Gong, J. Cartlidge, R. Bai, Y. Yue, Q. Li, and G. Qiu, “Extracting activity patterns from taxi trajectory data: a two-layer
framework using spatio-temporal clustering, Bayesian probability and Monte Carlo simulation,” International Journal of
Geographical Information Science, vol. 34, no. 6, pp. 1210–1234, 2020, doi: 10.1080/13658816.2019.1641715.
[17] Y. Yang, J. Cai, H. Yang, J. Zhang, and X. Zhao, “TAD: A trajectory clustering algorithm based on spatial-temporal density
analysis,” Expert Systems with Applications, vol. 139, Jan. 2020, doi: 10.1016/j.eswa.2019.112846.
[18] A. I. J. Tostes, F. D. L. P. Duarte-Figueiredo, R. Assunção, J. Salles, and A. A. F. Loureiro, “From data to knowledge,” in
Proceedings of the 2nd ACM SIGKDD International Workshop on Urban Computing, Aug. 2013, pp. 1–8, doi:
10.1145/2505821.2505831.
[19] A. Muñoz-Villamizar, E. L. Solano-Charris, M. AzadDisfany, and L. Reyes-Rubiano, “Study of urban-traffic congestion based on
Google Maps API: the case of Boston,” IFAC-PapersOnLine, vol. 54, no. 1, pp. 211–216, 2021, doi: 10.1016/j.ifacol.2021.08.079.
[20] G. Georgoulas, A. Konstantaras, E. Katsifarakis, C. D. Stylios, E. Maravelakis, and G. J. Vachtsevanos, “‘Seismic-mass’ density-
based algorithm for spatio-temporal clustering,” Expert Systems with Applications, vol. 40, no. 10, pp. 4183–4189, 2013, doi:
10.1016/j.eswa.2013.01.028.
[21] N. Nazia, J. Law, and Z. A. Butt, “Spatiotemporal clusters and the socioeconomic determinants of COVID-19 in Toronto

Int J Artif Intell, Vol. 14, No.1, February 2025: 397-407

Int J Artif Intell ISSN:2252-8938  407

neighbourhoods, Canada,” Spatial and Spatio-temporal Epidemiology, vol. 43, Nov. 2022, doi: 10.1016/j.sste.2022.100534.
[22] S. Deb and S. Karmakar, “A novel spatio-temporal clustering algorithm with applications on COVID-19 data from the United
States,” Computational Statistics and Data Analysis, vol. 188, 2023, doi: 10.1016/j.csda.2023.107810.
[23] J. Sodoge, C. Kuhlicke, and M. M. de Brito, “Automatized spatio-temporal detection of drought impacts from newspaper articles
using natural language processing and machine learning,” SSRN Electronic Journal, 2022, doi: 10.2139/ssrn.4178096.
[24] X. Liu and D. Lv, “Spatial and temporal characteristics, spatial clustering and governance strategies for regional development of
social enterprises in China,” Heliyon, vol. 10, no. 4, Feb. 2024, doi: 10.1016/j.heliyon.2024.e26246.
[25] F. Y. Foo, N. A. Rahman, F. Z. S. Abdullah, and N. S. A. Naeeim, “Spatio-temporal clustering analysis of COVID-19 cases in
Johor,” Infectious Disease Modelling, vol. 9, no. 2, pp. 387–396, 2024, doi: 10.1016/j.idm.2024.01.009.
[26] H. Peng, W. Li, C. Jin, H. Yang, and J. Guan, “MuSTC: A multi-stage spatio–temporal clustering method for uncovering the
regionality of global SST,” Atmosphere, vol. 14, no. 9, 2023, doi: 10.3390/atmos14091358.
[27] N. Madyavanhu, M. D. Shekede, S. Kusangaya, D. M. Pfukenyi, S. Chikerema, and I. Gwitira, “Bovine anaplasmosis in Zimbabwe:
spatio-temporal distribution and environmental drivers,” Veterinary Quarterly, vol. 44, no. 1, pp. 1–16, 2024, doi:
10.1080/01652176.2024.2306210.
[28] M. M. Nemukula, C. Sigauke, H. Chikoore, and A. Bere, “Modelling drought risk using bivariate spatial extremes: application to
the Limpopo Lowveld Region of South Africa,” Climate, vol. 11, no. 2, 2023, doi: 10.3390/cli11020046.
[29] S. Shivakumar et al., “Examining leopard attacks: spatio-temporal clustering of human injuries and deaths in Western Himalayas,
India,” Frontiers in Conservation Science, vol. 4, 2023, doi: 10.3389/fcosc.2023.1157067.
[30] R. Tang, G. Hou, and R. Du, “Isolated or colocated? exploring the spatio-temporal evolution pattern and influencing factors of the
attractiveness of residential areas to restaurants in the central Urban Area,” ISPRS International Journal of Geo-Information, vol.
12, no. 5, May 2023, doi: 10.3390/ijgi12050202.
[31] K. Tripathi, “The novel hierarchical clustering approach using self-organizing map with optimum dimension selection,” Health
Care Science, vol. 3, no. 2, pp. 88–100, 2024, doi: 10.1002/hcs2.90.
[32] T. Z. Nigussie, T. T. Zewotir, and E. K. Muluneh, “Detection of temporal, spatial and spatiotemporal clustering of malaria incidence
in northwest Ethiopia, 2012–2020,” Scientific Reports, vol. 12, no. 1, 2022, doi: 10.1038/s41598-022-07713-3.
[33] X. Yu et al., “Epidemiological characteristics and spatio-temporal analysis of brucellosis in Shandong Province, 2015–2021,” BMC
Infectious Diseases, vol. 23, no. 1, 2023, doi: 10.1186/s12879-023-08503-6.
[34] National Center for Seismology, “Seismological data: earthquake catalogue, Aug. 2019 to Jan. 2024,” National Center for
Seismology, Ministry of Earth Sciences, Government of India. [Online]. Available:
https://riseq.seismo.gov.in/riseq/earthquake/archive

BIOGRAPHIES OF AUTHORS

Ms. Swati Meshram is a Research Scholar in Computer Science and Engineering,

has Master of Engineering degree in computer engineering and is working as an Assistant
Professor, P.G. Department of Computer Science, SNDT Women’s University, Maharashtra,
India. Her areas of interest are data mining, machine learning, and databases. She can be
contacted at email: [email protected].

Dr. Kishor P. Wagh is a Research Supervisor and working as an Assistant

Professor at Department of Information Technology, Government College of Engineering,
Amravati. He has more than 20 years of experience in teaching and research in the field of
computer science and engineering and information technology. His area of interest is in data
mining, machine learning, databases, and object oriented methodology. He can be contacted at
email: [email protected].

Dynamic spatio-temporal pattern discovery: a novel grid and density-based … (Swati Meshram)