University of Science and Technology of Hanoi

Analysis of Spatial & Temporal Data

Lecturer: Nguyen Xuan Thanh

Spatio - Temporal Dynamics of

Urban Air Pollution

Student Name Student ID

Nguyen Chi Quang 22BA13261

Nguyen Dinh Lien Thanh 22BA13288
Nguyen Hoang Tung 22BA13388
Ngo Thanh Dat 23BI14090
Pham Cong Duyet 23BI14136
Duong Tuan Kiet 22BI13230

Hanoi, February 2026

 Abstract

Urban air pollution changes quickly over time and can also spread across nearby
cities. This study analyzes the spatio-temporal dynamics of urban air pollution and
builds an Early Warning System (EWS) to answer a practical question: Will tomorrow
be an extreme pollution day? Hourly data are collected from the AQICN API (August
2025–January 2026) for 28 cities and aggregated to daily series. Extreme events are de-
fined in a data-driven way using a city-specific 80th percentile threshold of daily AQI.
The analysis combines statistical profiling of AQI distributions, temporal methods (lag
dependence, autocorrelation, and wavelet analysis), and spatial/spatio-temporal meth-
ods (Global Moran’s I, LISA hotspots, and variograms). Weather–pollution regimes are
also identified using DBSCAN on standardized meteorology and gaseous-pollutant vari-
ables. Finally, logistic regression models are evaluated to predict extreme tomorrow.
Results show that adding continuous meteorology/chemistry drivers improves prediction
compared with using temporal features alone, while regime labels mainly support in-
terpretation. Overall, the framework provides a clear and interpretable approach for
understanding and forecasting extreme pollution risk.

2
Contents

1 Introduction 6
1.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.2 Problem Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.3 Project Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.4 Report Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

2 Methodology 8
2.1 Data and Study Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.1.1 Data source and time range . . . . . . . . . . . . . . . . . . . . . 8
2.1.2 Study cities (core vs extended) and variables . . . . . . . . . . . . 8
2.1.3 Train/test split and standardization . . . . . . . . . . . . . . . . . 9
2.2 Preprocessing and Feature Engineering . . . . . . . . . . . . . . . . . . . 9
2.2.1 Cleaning and aggregation (hourly to daily) . . . . . . . . . . . . . 9
2.2.2 Extreme definition and target construction . . . . . . . . . . . . . 9
2.2.3 Temporal features . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.2.4 Multivariate features . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.3 Analytical Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.3.1 Statistical profiling . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.3.2 Temporal methods . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.3.3 Spatial Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.3.4 Spatio-Temporal Analysis . . . . . . . . . . . . . . . . . . . . . . 12
2.3.5 Regime discovery . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.3.6 Predictive models . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

3 Results 14
3.1 Statistical Profiling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.2 Temporal Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.2.1 Temporal AQI Dynamics . . . . . . . . . . . . . . . . . . . . . . . 17
3.2.2 Lag Dependency - Temporal Persistence . . . . . . . . . . . . . . 18
3.2.3 Multi-Scale Temporal Dynamics . . . . . . . . . . . . . . . . . . . 20
3.3 Spatial Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.3.1 Step 1: Testing for Spatial Autocorrelation . . . . . . . . . . . . . 22
3.3.2 Step 2: Local Hotspot Identification . . . . . . . . . . . . . . . . . 23
3.3.3 Step 3: Estimating Spatial Range . . . . . . . . . . . . . . . . . . 24
3.3.4 Synthesis of Spatial Structure . . . . . . . . . . . . . . . . . . . . 25
3.4 Spatio-Temporal Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.4.1 Step 1: Temporal Evolution of Spatial Autocorrelation . . . . . . 26
3.4.2 Step 2: Spatial Stability Analysis . . . . . . . . . . . . . . . . . . 27
3.4.3 Step 3: Seasonal Variogram Analysis (The Heterogeneity Paradox) 28

3
 CONTENTS

3.5 Multivariate Regimes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30

3.5.1 Multicollinearity Check (VIF) . . . . . . . . . . . . . . . . . . . . 30
3.5.2 Discover regimes with DBSCAN . . . . . . . . . . . . . . . . . . . 30
3.5.3 EWS evidence: do regimes change next-day extreme risk? . . . . 31
3.6 Predictive Validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.6.1 Discrimination performance (ROC-AUC and PR-AUC) . . . . . . 32
3.6.2 Operational performance at a fixed threshold . . . . . . . . . . . . 33

4 Conclusion and Future Work 34

4.1 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
4.2 Limitations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
4.3 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34

4
List of Figures

3.1 Evidence of right-skewed AQI distributions. The mean–median gap across

cities indicates that extreme pollution events pull the mean toward the
right tail. The Beijing example illustrates how the right tail increases the
mean above the median. . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.2 Distribution fitting results across the eight cities. The plot shows the
∆AIC between the best and the second-best model. Larger values indicate
stronger statistical evidence for the winning distribution. . . . . . . . . . 15
3.3 City-specific 80th percentile thresholds for extreme pollution events. The
red vertical line indicates the AQI level used to define extreme pollution
days in each city. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.4 Temporal AQI dynamics for (Beijing, Hamburg, and Seoul) . . . . . . . . 17
3.5 Lag relationship between the current AQI and the previous day’s AQI for
selected cities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
3.6 Wavelet power spectrum of AQI for Beijing . . . . . . . . . . . . . . . . . 20
3.7 Time series of Global Moran’s I showing consistently positive values across
all months. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.8 LISA Cluster Map: Red dots represent High-High Hotspots, Blue dots
represent Low-Low Coldspots, and Grey dots represent not significant areas. 23
3.9 Spatial Range: The Variogram curve plateaus at 5250 km, showing the
maximum distance of spatial correlation. . . . . . . . . . . . . . . . . . . 24
3.10 Seasonal Pattern: Moran’s I spikes dramatically during the Winter months,
proving seasonal elasticity. . . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.11 Seasonal Range Elasticity: The Winter curve reaches its Sill much faster
due to intense spatial heterogeneity. . . . . . . . . . . . . . . . . . . . . . 28
3.12 kNN distance plot (k = 5) used to select the DBSCAN eps. . . . . . . . 30
3.13 DBSCAN clustering result for core cities (eps=1.5, minPts=5). . . . . . 31
3.14 ROC curves on the test set for Models A–C. Model C dominates most of
the curve, consistent with its higher ROC-AUC. . . . . . . . . . . . . . . 33

5
Chapter 1

Introduction

1.1 Background
Urban air pollution is one of the most serious environmental challenges in modern
cities. Rapid urbanization, industrial development, and increasing transportation activ-
ities have contributed to the deterioration of air quality in many metropolitan regions.
High concentrations of pollutants such as particulate matter (PM) and nitrogen oxides
(NO2 ) pose serious risks to public health, environmental sustainability, and urban living
conditions.
Air pollution is also highly dynamic across both space and time. Pollution levels can
fluctuate from day to day due to changing meteorological conditions, human activities,
and atmospheric transport processes. Pollutants may also travel across long distances and
affect neighboring cities, creating regional pollution patterns that are difficult to capture
with simple descriptive analyses. Therefore, beyond describing pollution patterns, an
operational goal is to forecast short-term extreme-risk days to support early warnings.

1.2 Problem Statement

Despite the growing availability of environmental data, predicting extreme air pol-
lution events remains challenging. Urban pollution is influenced by multiple interacting
factors, including distributional properties, short-term temporal dependence, spatial clus-
tering across cities, and meteorological–chemical conditions. These drivers can combine
to create context-dependent patterns that are not well explained when temporal, spatial,
and atmospheric factors are analyzed separately.
A second challenge is defining what constitutes a high-risk pollution event. In practice,
warning thresholds are often chosen in an ad hoc way rather than being grounded in
statistical evidence. Without a data-driven definition of extremes, early warning systems
may fail to represent real environmental risk.
This project focuses on a single operational question:

Will tomorrow be an extreme pollution day?

Answering this question requires an integrated framework that links statistical profiling,
temporal memory, spatial structure, and multivariate atmospheric conditions.

6
 CHAPTER 1. INTRODUCTION

1.3 Project Objectives

The main objective of this project is to develop an integrated analytical framework
to understand urban air pollution dynamics and to support an early warning system
(EWS) for next-day extreme events. To achieve this objective, the study profiles city-
level pollution distributions and defines extreme events using a data-driven, city-specific
threshold based on daily aggregated AQI; engineers temporal indicators of persistence
and momentum to capture short-term memory; quantifies spatial and spatio-temporal
structure using global and local clustering measures and spatial-range diagnostics; discov-
ers multivariate weather–pollution regimes using unsupervised learning on standardized
meteorology and chemistry features and assesses whether regime context is associated
with elevated next-day extreme risk; and validates predictive performance using logistic
regression models with out-of-sample evaluation, reporting ROC-AUC, PR-AUC, and
operational threshold-based confusion metrics.

1.4 Report Structure

This report is organized into analytical modules that progressively build a framework
for understanding and predicting urban air pollution dynamics.
Module 1 performs statistical profiling and defines a data-driven threshold for extreme
events. Module 2 examines temporal dynamics and constructs persistence/momentum
features for early warning. Modules 3 and 4 investigate spatial and spatio-temporal
interactions, including clustering, hotspot structure, and seasonal variation.
Module 5 discovers multivariate weather–pollution regimes using DBSCAN and eval-
uates whether regimes are associated with higher next-day extreme risk. Finally, Module
6 compares multiple logistic-regression EWS specifications to test whether regime context
or continuous meteorology drivers improve next-day extreme-event prediction.

7
Chapter 2

Methodology

2.1 Data and Study Design

2.1.1 Data source and time range
Hourly air-quality observations were collected from the AQICN platform ([Link]
[1]) via its public API. The raw dataset contains 95,810 rows and 16 columns, covering
August 2025 to January 2026. It includes hourly AQI (aqi hour), major pollutants
(PM2.5 , PM10 , NO2 , SO2 , CO, O3 ), key meteorological variables (temperature, humidity,
pressure, wind speed), and spatial metadata (city and latitude/longitude).

2.1.2 Study cities (core vs extended) and variables

To support both robust spatio-temporal analysis and broader regional context, the
study cities are split into two groups. The Core group (deep-dive analysis) contains
eight cities representing diverse geographic regions and pollution patterns: Beijing, Seoul,
Fukuoka, Zhongshan (Taiwan), Warsaw, Nagoya, Hamburg, and Kawasaki. The Ex-
tended group provides broader spatial coverage and includes 20 cities. Spatial and
spatio-temporal modules use the Extended set to capture broader regional context, while
predictive EWS validation is restricted to the Core cities due to data-quality constraints
and the need for consistent daily time series.
Core cities are selected using strict data-quality constraints. Each core city must
have at least 120 days of coverage, hourly completeness (observed/expected hours) of
at least 0.80, and no severe discontinuity in the time series. A time series is considered
discontinuous if the maximum gap exceeds 24 hours, unless the total number of missing
hours in long gaps (gaps >6h) remains below 5% of expected hours. If more cities satisfy
these constraints than needed, priority is given to cities with higher completeness, longer
coverage, stable missingness across months, and greater geographic diversity. Table 2.1
summarizes the variables used in this study.

8
 CHAPTER 2. METHODOLOGY

Table 2.1: Summary of variables in the collected dataset

Category Variable Name Description

Target Variable aqi hour Hourly Air Quality Index
dominantpol aqi Dominant pollutant of the hour
Pollutants pm25 Particulate matter ≤ 2.5µm
pm10 Particulate matter ≤ 10µm
no2 Nitrogen dioxide
so2 Sulfur dioxide
co Carbon monoxide
o3 Ozone
Meteorology temperature Ambient temperature
humidity Relative humidity
pressure Atmospheric pressure
wind speed Wind speed
Spatial Context lat, lon Latitude and Longitude coordinates
city Observation location

2.1.3 Train/test split and standardization

For predictive validation, a time-aware split is applied to mimic real forecasting:
the first 80% of the time-ordered observations are used for training and the remaining
20% for testing. Continuous meteorology/chemistry predictors are standardized using
training-set means and standard deviations only, and the same scaling is applied to the
test set to prevent information leakage.

2.2 Preprocessing and Feature Engineering

2.2.1 Cleaning and aggregation (hourly to daily)
The analysis uses the cleaned AQICN dataset as the main input for all modules. After
cleaning, the dataset contains 80,895 rows and 16 columns across 28 cities. Because
environmental sensor data may contain missing values and occasional anomalies, data
types are standardized (timestamps, numeric pollutant and meteorological variables, and
city identifiers) and invalid records are removed. For short missing segments within the
time series, linear interpolation is applied using adjacent hours to preserve the temporal
structure without introducing artificial jumps. For modules that require daily resolution,
hourly data are aggregated to a city–day level using simple summary statistics, such as
daily means for meteorology and pollutant concentrations. For EWS labeling, the daily
AQI series is computed from hourly AQI using a robust statistic (daily median) to reduce
the influence of short-lived hourly outliers.

2.2.2 Extreme definition and target construction

Extreme pollution events are defined in a data-driven and city-specific way. For each
core city, the 80th percentile of the daily AQI distribution is computed and a day is
labeled as an extreme event if its daily AQI exceeds this threshold. To match the early
warning objective, the prediction target extreme tomorrow is created by shifting the

9
 CHAPTER 2. METHODOLOGY

label forward by one day within each city. The resulting task is a binary classification
problem: predicting whether tomorrow will be an extreme day.

2.2.3 Temporal features

To capture short-term persistence and momentum, temporal predictors are engineered
from the daily AQI series. The main features are Lag1, defined as yesterday’s daily AQI,
and Delta AQI, defined as today’s AQI minus yesterday’s AQI. In addition, extreme lag1
is created as a binary indicator of whether yesterday was extreme. These features form
the temporal baseline for the logistic-regression EWS.

2.2.4 Multivariate features

To represent atmospheric and chemical drivers, daily multivariate features are con-
structed from hourly measurements. For each city and day, daily means of temperature,
humidity, pressure, wind speed, NO2 , SO2 , CO, and O3 are computed. These variables
are standardized (z-scored) before multivariate analysis to place them on a comparable
scale. They are then used for regime discovery (DBSCAN) and as continuous predictors
in the predictive validation models.
The variables dominantpol aqi, PM2.5 , and PM10 are excluded from the driver set be-
cause they are not suitable as upstream predictors in the EWS design. dominantpol aqi
is a derived label that summarizes which pollutant dominates the AQI, while PM2.5 and
PM10 largely reflect the pollution outcome and are often strongly aligned with aqi hour.
To keep the regime analysis and forecasting step interpretable and not dominated by
the AQI definition, the driver set focuses on meteorology and gaseous pollutants (NO2 ,
SO2 , CO, O3 ), which better represent atmospheric conditions that can precede next-day
extremes.

2.3 Analytical Methods

2.3.1 Statistical profiling
The analysis is based on real-world air quality data collected from the AQICN API.
The dataset contains hourly AQI observations from multiple monitoring stations. For
statistical profiling, the study focuses on eight core cities representing different geographic
regions and pollution patterns: Beijing, Seoul, Fukuoka, Zhongshan, Warsaw, Nagoya,
Hamburg, and Kawasaki.
Hourly AQI observations were aggregated into daily AQI values to create consistent
time series for distribution analysis. Four candidate probability distributions were tested:
Normal, Lognormal, Gamma, and Weibull. Model parameters were estimated using
maximum likelihood estimation (MLE).
The goodness-of-fit of each model was evaluated using the Akaike Information Crite-
rion (AIC), the Bayesian Information Criterion (BIC), and the Anderson–Darling (AD)
test. Lower AIC and BIC values indicate better model performance, while the AD test
provides additional sensitivity to deviations in the distribution tails.

10
 CHAPTER 2. METHODOLOGY

2.3.2 Temporal methods

To characterize the temporal dynamics of air pollution and extract meaningful features
for subsequent analysis, a time-series analytical framework is applied to the daily Air
Quality Index (AQI). Let {AQIt }Tt=1 denote the AQI time series for a given city over T
days. The temporal analysis focuses primarily on evaluating short-term persistence and
identifying multi-scale oscillatory patterns within the pollution data.

Lag Dependency and Autocorrelation To determine whether current air pollution

levels are influenced by recent conditions, the temporal persistence of the AQI series is
examined. The relationship between consecutive days is first evaluated using lag analysis
(comparing AQIt with AQIt−1 ). To formally quantify this dependence, the first-lag
autocorrelation coefficient (ACF (1)) is calculated:

t=2 (AQIt − µ)(AQIt−1

PT
− µ)
ACF (1) =
t=1 (AQIt − µ)
PT 2

where µ is the mean of the AQI time series. The ACF (1) provides a statistical measure
of how strongly pollution levels persist from one day to the next, with higher values
indicating stronger temporal dependence.

Multi-Scale Temporal Dynamics While lag analysis and autocorrelation assess im-
mediate, day-to-day serial dependence, air pollution dynamics are often driven by complex
processes that operate across broader and varying time scales, from short-term weather
cycles to seasonal trends. To capture these multi-scale variations, wavelet analysis is
applied to the AQI time series. Unlike traditional spectral methods that assume station-
arity, wavelet analysis transforms the one-dimensional time series into a two-dimensional
time-frequency representation, allowing for the detection of localized oscillatory behav-
iors.
In this study, the wavelet power spectrum is generated to visualize the intensity
of these temporal oscillations across different frequencies. To summarize this complex
information into a single comparable feature across cities, the mean wavelet energy is
calculated. This metric quantifies the overall strength of multi-scale temporal fluctuations
in the AQI series.

11
 CHAPTER 2. METHODOLOGY

2.3.3 Spatial Analysis

Air pollution is rarely a purely local issue. To uncover the physical structure of
pollution—whether it spreads randomly or forms connected networks—a sequence of
geostatistical methods is applied to quantify spatial relationships across the studied cities.
To determine whether polluted cities exhibit spatial clustering or operate indepen-
dently, the Global Moran’s I statistic is utilized. This method measures spatial auto-
correlation by evaluating a city’s air quality against its geographical neighbors using a
k-Nearest Neighbors (k-NN) spatial weight matrix. The k-NN approach is specifically
chosen over distance-based weights to ensure all cities have an equal number of neigh-
bors, which is crucial for balancing our sparsely distributed global dataset. The resulting
statistic ranges from -1 to +1, where values near zero indicate a completely random dis-
tribution, and positive values (particularly above 0.30) signify moderate to strong spatial
clustering.
While the global statistic confirms the general presence of clustering, Local Indicators
of Spatial Association (LISA) are applied to locate specific pollution epicenters. LISA de-
composes the global Moran’s I to compare individual cities with their local surroundings.
At a significance level of p < 0.05, this method categorizes cities into four spatial quad-
rants, most notably identifying “High-High” hotspots—heavily polluted cities surrounded
by other polluted areas—which serve as primary targets for intervention.
Finally, to quantify the physical extent of these pollution clusters, empirical vari-
ograms are constructed and fitted with a Spherical model. The variogram maps the
relationship between the geographic distance separating two cities and the difference in
their pollution levels (semivariance). The Spherical model is highly robust for sparse
spatial datasets and provides a definitive correlation cutoff. The fitted model extracts
three critical parameters: the nugget (micro-scale variation and sensor noise), the sill
(maximum spatial variance), and the range. The range represents the exact distance at
which spatial correlation drops to zero, thereby defining the geographic boundary of a
hotspot’s influence.

2.3.4 Spatio-Temporal Analysis

Building upon the static spatial structure, spatio-temporal analysis is conducted to
evaluate the seasonal elasticity of the pollution network. It addresses whether the spatial
architecture of air pollution remains rigid year-round or fluctuates with meteorological
shifts.
First, the temporal evolution of spatial autocorrelation is tracked by calculating the
Global Moran’s I score for each month and grouping the results by season. This deter-
mines whether the strength of regional clustering amplifies during specific periods, such
as winter temperature inversions. Subsequently, a spatial shift analysis is performed by
comparing the seasonal LISA cluster classifications. This step calculates a stability index
to ascertain whether pollution epicenters migrate across regions or remain geographically
anchored regardless of the weather.
Finally, seasonal variogram models are fitted to compare the spatial ranges and sills
between summer and winter. This mathematical comparison aims to reveal whether
seasonal atmospheric stability expands the physical reach of pollution globally or hyper-
concentrates it within existing regional hotspots. Understanding this seasonal hetero-
geneity paradox is essential for configuring dynamic, season-specific confidence intervals
in the final Early Warning System.

12
 CHAPTER 2. METHODOLOGY

2.3.5 Regime discovery

To capture typical combinations of weather and pollutant conditions, daily weather–
pollution regimes are discovered using DBSCAN clustering. An 8-dimensional feature
vector is constructed for each city–day from standardized daily means of temperature,
humidity, pressure, wind speed, NO2 , SO2 , CO, and O3 . DBSCAN is selected because it
can identify a dense “baseline” pattern while labeling rare days as noise (outliers), which
is useful for early warning. The parameter minPts is set to 5, and eps is selected using
the kNN-distance elbow method.

2.3.6 Predictive models

The Early Warning System (EWS) is validated using logistic regression to predict the
binary target extreme tomorrow. To isolate the value of different information sources,
three model specifications are compared using the same time-aware training set and the
same held-out test set. Model A is a temporal baseline with city fixed effects, a time
index, and short-term memory features (Lag1, Delta AQI, and extreme lag1). Model
B adds the Weather Regime label to test whether regime context improves prediction.
Model C replaces the regime label with standardized continuous meteorology/chemistry
drivers (the same 8 variables used for DBSCAN). Performance is reported using test
ROC-AUC and PR-AUC. Operational performance at a fixed threshold (τ = 0.45) is
summarized using precision, recall, F1-score, and warning rate.

13
Chapter 3

Results

3.1 Statistical Profiling

The objective of this analysis is to examine the statistical distribution of daily Air
Quality Index (AQI) values and to test whether the data follow a Gaussian (Normal)
distribution. Environmental pollution data often contain extreme events, which may lead
to asymmetric and heavy-tailed distributions. By analyzing the empirical distribution of
AQI, we evaluate its skewness and tail behavior.
As shown in Figure 3.1, AQI distributions exhibit clear right-skewness across the
studied cities. The positive gap between the mean and the median indicates that ex-
treme pollution events pull the average AQI toward the right tail. The Beijing example
illustrates this effect clearly: while most observations are concentrated at moderate lev-
els, a smaller number of very high AQI values extend the right tail of the distribution.
This pattern suggests that AQI data are not symmetric and therefore may not follow a
Gaussian distribution.

Figure 3.1: Evidence of right-skewed AQI distributions. The mean–median gap across
cities indicates that extreme pollution events pull the mean toward the right tail. The
Beijing example illustrates how the right tail increases the mean above the median.

The distribution fitting results indicate that the Normal distribution generally pro-
vides the poorest fit to the AQI data. In most cities, the Normal model produces sig-
nificantly higher AIC and BIC values compared with the other candidate distributions.

14
 CHAPTER 3. RESULTS

This confirms that the assumption of symmetric Gaussian behavior does not accurately
describe the statistical structure of daily AQI.
In contrast, the Gamma and Lognormal distributions consistently provide better fits.
These models capture the asymmetric structure of AQI data, where most observations
occur at moderate pollution levels while a smaller number of observations extend toward
very high values.
Figure 3.2 summarizes the distribution fitting comparison across the eight cities. The
horizontal axis shows the ∆AIC between the best and the second-best models. Larger
values indicate stronger statistical evidence supporting the winning distribution.

Figure 3.2: Distribution fitting results across the eight cities. The plot shows the ∆AIC
between the best and the second-best model. Larger values indicate stronger statistical
evidence for the winning distribution.

The analysis demonstrates that AQI does not follow a Gaussian distribution but
instead shows right-skewed and heavy-tailed characteristics. Because of this behavior,
extreme pollution events are defined using a percentile-based threshold rather than a
fixed AQI value.
Specifically, the 80th percentile of the AQI distribution is used as the threshold for
extreme pollution events in each city. Figure 3.3 illustrates these city-specific thresholds.
The red vertical line in each histogram indicates the AQI level that separates typical days
from extreme pollution days.

15
 CHAPTER 3. RESULTS

Figure 3.3: City-specific 80th percentile thresholds for extreme pollution events. The red
vertical line indicates the AQI level used to define extreme pollution days in each city.

The resulting extreme pollution indicator is later used as the binary target variable
for the Early Warning System.

3.2 Temporal Dynamics

The goal is to analyze the temporal behavior of air quality and extract meaningful
features from the AQI time series. Air pollution exhibits strong temporal variability, and
understanding these patterns is important for characterizing pollution dynamics across
cities and understanding these temporal patterns is important for identifying pollution
dynamics.
This module focuses on examining how AQI values evolve across time and whether
current pollution levels are influenced by previous observations. To achieve this, several
analytical techniques are applied, including time-series visualization, lag analysis, short-
term change measurement, and wavelet-based frequency analysis. These techniques help
reveal important temporal characteristics such as persistence, variability, and multi-scale
fluctuations in air pollution levels.
The extracted temporal indicators provide quantitative information about how pol-
lution behaves over time in different cities. These features are then used to support
further analysis and serve as useful inputs for subsequent modules in the overall analyti-
cal pipeline.

16
 CHAPTER 3. RESULTS

3.2.1 Temporal AQI Dynamics

To understand the overall behavior of air pollution across time, the daily AQI time
series of several representative cities are first visualized. This step provides an initial
view of how pollution levels fluctuate over time and helps identify trends, spikes, and
variability in air quality.

Figure 3.4: Temporal AQI dynamics for (Beijing, Hamburg, and Seoul)

From Figure 3.4, clear differences in temporal pollution patterns can be observed
between the cities. Beijing exhibits the highest variability with frequent and intense
spikes, indicating unstable pollution levels during the observed period. Seoul also shows
noticeable fluctuations with several peaks during the later months, suggesting periods of
increased pollution. In contrast, Hamburg presents relatively lower AQI levels and more
moderate variations, reflecting generally cleaner air conditions compared to the other
cities.
To complement the visual analysis, Table 1 summarizes the average AQI level for each
city during the study period.

Table 3.1: Average AQI levels across selected cities

City Median AQI

Beijing 72.0
Fukuoka 42.0
Hamburg 27.5
Kawasaki 38.0
Nagoya 46.0
Seoul 54.5
Warsaw (Poland) 51.0
Zhongshan (Taiwan) 42.0

The summary statistics in Table 3.1 reveal clear differences in air pollution levels
among the selected cities. Beijing records the highest median AQI value (72.0), indicating

17
 CHAPTER 3. RESULTS

more severe pollution conditions compared to the other locations. In contrast, Hamburg
shows the lowest median AQI value (27.5), suggesting relatively cleaner air conditions.
Several cities fall within a moderate pollution range. Seoul (54.5) and Warsaw (51.0)
display moderately elevated AQI levels, while Nagoya (46.0) remains slightly lower but
still within the mid-range. Meanwhile, Fukuoka (42.0) and Zhongshan (42.0) present sim-
ilar median AQI values, indicating relatively moderate air quality conditions. Kawasaki
(38.0) shows somewhat lower pollution levels compared with most Asian cities in the
sample.
Overall, the results highlight substantial variation in air pollution levels across the
studied cities. Beijing clearly stands out as the most polluted city in terms of median
AQI, while Hamburg consistently exhibits the cleanest air conditions. These differences
provide an important baseline for further analysis of the temporal dynamics of AQI across
locations.
While the time-series visualization provides an overall view of how AQI levels fluctuate
over time, it does not explicitly quantify the relationship between consecutive observa-
tions. To better understand whether current pollution levels are influenced by recent
conditions, the next analysis examines the lag dependency in the AQI series.

3.2.2 Lag Dependency - Temporal Persistence

To investigate the temporal persistence of air pollution, both the lagged relationship
between consecutive AQI values and the first-lag autocorrelation (ACF1) are examined.
These indicators help evaluate the extent to which current air pollution levels are influ-
enced by previous observations.

Figure 3.5: Lag relationship between the current AQI and the previous day’s AQI for
selected cities

The scatter plot illustrates a positive relationship between the AQI of a given day and
the AQI of the previous day across the selected cities. In most cases, higher AQI values
tend to follow periods with already elevated pollution levels, indicating that air pollution
conditions often persist across consecutive days.

18
 CHAPTER 3. RESULTS

Table 3.2: Temporal persistence indicators across cities

City Median Lag1 AQI ACF1

Beijing 71.50 0.3997
Fukuoka 42.00 0.3898
Hamburg 27.75 0.5711
Kawasaki 38.00 0.2864
Nagoya 46.25 0.3179
Seoul 53.75 0.6108
Warsaw (Poland) 51.00 0.7209
Zhongshan (Taiwan) 42.00 0.6220

Table 3.2 summarizes the temporal persistence indicators of AQI across the selected
cities. While the mean Lag1 AQI reflects the general level of pollution carried over from
the previous day, the ACF1 statistic directly measures the strength of serial dependence
in the time series.
Cities such as Warsaw (0.7209), Zhongshan (0.6220), and Seoul (0.6108) exhibit rel-
atively high autocorrelation values, indicating strong persistence in daily air pollution
levels. Hamburg (0.5711) also shows a noticeable degree of temporal dependence. In
contrast, Kawasaki (0.2864) and Nagoya (0.3179) display weaker autocorrelation values,
suggesting more variable day-to-day pollution dynamics. Beijing (0.3997) and Fukuoka
(0.3898) fall within a moderate range of temporal dependence.
Overall, the combined evidence from Lag1 AQI and ACF1 indicators suggests that
air pollution levels tend to exhibit temporal persistence across most cities, although the
strength of this dependence varies across urban environments.
While lag dependency and autocorrelation analysis reveal the degree of short-term
persistence in AQI values, they mainly capture relationships between consecutive obser-
vations. Such measures provide limited insight into broader temporal structures that may
emerge across longer time horizons. To further investigate these multi-scale temporal dy-
namics and identify oscillatory patterns occurring at different time frequencies, wavelet
analysis is applied to the AQI time series.

19
 CHAPTER 3. RESULTS

3.2.3 Multi-Scale Temporal Dynamics

Figure 3.6 presents the wavelet power spectrum of the AQI time series, which reveals
temporal oscillations across multiple time scales. Wavelet analysis is useful for captur-
ing complex patterns in air pollution dynamics that may occur at different temporal
frequencies.
Beijing is used as a representative example due to its higher pollution variability.

Figure 3.6: Wavelet power spectrum of AQI for Beijing

The wavelet power spectrum for Beijing shows several high-energy regions, particu-
larly at shorter periods between roughly 4 and 12 days. This indicates strong short-term
oscillations in AQI levels. Overall, the pattern suggests pronounced multi-scale variability
in Beijing’s air pollution dynamics.

Table 3.3: Wavelet energy summary across cities

City Mean Wavelet Energy

Beijing 0.1400
Fukuoka 0.1251
Hamburg 0.1060
Kawasaki 0.1242
Nagoya 0.1260
Seoul 0.1010
Warsaw (Poland) 0.0992
Zhongshan (Taiwan) 0.1085

Table 3.3 summarizes the average wavelet energy for each city, reflecting the strength
of multi-scale temporal oscillations in AQI. Beijing shows the highest mean wavelet energy
(0.1400), indicating stronger temporal variability across different time scales. Nagoya
(0.1260), Fukuoka (0.1251), and Kawasaki (0.1242) also exhibit relatively high energy
levels, suggesting noticeable oscillatory patterns. In contrast, Seoul (0.1010) and Warsaw
(0.0992) present lower values, indicating weaker multi-scale fluctuations.

20
 CHAPTER 3. RESULTS

Overall, the results suggest that the strength of temporal oscillations in AQI varies
across cities, with some urban environments exhibiting more pronounced multi-scale dy-
namics than others.

Conclusion
This section investigated the temporal dynamics of air quality across multiple cities
using several complementary time-series analysis techniques. The results show that AQI
levels exhibit clear temporal structure, including persistence across consecutive days,
short-term fluctuations in pollution levels, and multi-scale oscillatory behavior.
Lag analysis and autocorrelation results indicate that air pollution often persists from
one day to the next, although the strength of this dependence varies between cities. The
analysis of daily AQI differences further reveals that short-term variability differs across
locations, with some cities showing more volatile pollution dynamics. In addition, wavelet
analysis highlights the presence of multi-scale temporal oscillations, suggesting that air
pollution patterns are influenced by processes operating at different time scales.
Overall, these findings provide a comprehensive characterization of the temporal be-
havior of AQI and allow the extraction of several informative temporal indicators. These
indicators serve as useful quantitative features for the subsequent stages of the analyt-
ical framework, where they can support deeper comparative analysis and data-driven
modeling of urban air pollution patterns.

21
 CHAPTER 3. RESULTS

3.3 Spatial Analysis

Air pollution is rarely a purely local issue. This section uses geostatistical methods
to uncover the physical structure of pollution: does it spread randomly, or does it form
connected networks?

3.3.1 Step 1: Testing for Spatial Autocorrelation

Key Question: Do polluted cities stand alone, or do they group together with other
polluted neighbors?
To answer this, we calculated the Global Moran’s I statistic over six months. Moran’s I
measures whether cities with similar pollution levels cluster together in geographic space.
If the value is close to 0, pollution is random. If it is positive (especially > 0.30 in our
context), it indicates that strong clustering exists.

Figure 3.7: Time series of Global Moran’s I showing consistently positive values across
all months.

As shown in Figure 3.7, the Moran’s I value remains consistently positive (averaging
around 0.38) and never drops to zero. Consequently, air pollution is absolutely not
a random event. It operates as a persistently connected network. A dirty city will
inevitably pull its neighbors into a pollution cluster.

22
 CHAPTER 3. RESULTS

3.3.2 Step 2: Local Hotspot Identification

Key Question: Since pollution clusters exist, where exactly are the “hotspots” that
produce the most extreme pollution?
While Global Moran’s I proves that clustering happens, it does not tell us where.
Therefore, we used the LISA (Local Indicators of Spatial Association) algorithm to map
out specific cities and categorize them.

Figure 3.8: LISA Cluster Map: Red dots represent High-High Hotspots, Blue dots rep-
resent Low-Low Coldspots, and Grey dots represent not significant areas.

Figure 3.8 reveals a stark geographic reality. The “High-High” hotspots (red dots)
are densely concentrated in South Asia (especially Delhi, Mumbai, etc.). These cities are
not only highly polluted themselves but are surrounded by other heavily polluted areas.
Environmental responsibility is not shared equally. South Asia acts as the
“pollution engine” of the region. Any predictive model or policy must treat these anchored
hotspots as the primary risk sources.

23
 CHAPTER 3. RESULTS

3.3.3 Step 3: Estimating Spatial Range

Key Question: How far does a pollution hotspot’s influence travel? At what distance
is a city finally safe from a neighboring city’s bad air?
To find the physical size of this “pollution blanket,” we used a Spherical Variogram
model. The Variogram calculates the Range—the exact distance in kilometers at which
two cities stop influencing each other’s air quality.

Figure 3.9: Spatial Range: The Variogram curve plateaus at 5250 km, showing the
maximum distance of spatial correlation.

As shown in Figure 3.9 and our computational output, the model returned highly
specific parameters quantifying the spatial structure of pollution:

• Nugget (0.00): Zero micro-scale variation (no major sensor noise or hyperlocal
interruptions).

• Total Sill (2619.43): The maximum spatial variance (the plateau where correla-
tion ends).

• Range (5250 km): The exact distance where correlation drops to zero based on
the Spherical model cutoff.

The “blast radius” of air pollution is massive. Spatial correlation extends to exactly
5250 km—beyond this distance, cities become statistically independent. This Range pro-
vides a quantitative boundary for setting up regional environmental treaties and defining
the radar scope for our Early Warning System.
Note: With 28 globally distributed cities, this 5250 km range reflects large-scale re-
gional spatial structure rather than local atmospheric dispersion.

24
 CHAPTER 3. RESULTS

3.3.4 Synthesis of Spatial Structure

This section establishes three structural pillars of global air pollution:

1. Persistent spatial clustering: Global Moran’s I consistently stays above 0.30.

2. Anchored epicenters: LISA hotspots are heavily concentrated in South and East
Asia.

3. Massive correlation range: An effective spatial reach of 5250 km.

We have now mapped the spatial architecture of pollution. The analytical question
shifts: Does this rigid spatial architecture stay still, or does it flex and bend
with the seasons?

25
 CHAPTER 3. RESULTS

3.4 Spatio-Temporal Analysis

The previous section proved that air pollution has a massive spatial architecture. But
the Earth rotates, and seasons change. Does this pollution network stay still, or does
it “breathe” with the weather? This section analyzes how time and meteorology bend
spatial geography.

3.4.1 Step 1: Temporal Evolution of Spatial Autocorrelation

Key Question: Does the strength of the pollution network change depending on the
season?
To find out, we tracked the Global Moran’s I score over time, grouping the data by
season to test for temporal stability.

Figure 3.10: Seasonal Pattern: Moran’s I spikes dramatically during the Winter months,
proving seasonal elasticity.

The data reveals a massive Winter Amplification effect. During summer, the
spatial clustering is moderate. However, as winter approaches, the Moran’s I score jumps
by nearly 39.2%.
This indicates that winter does not merely make the air colder; it forces cities to bind
together. Temperature inversions (a common winter phenomenon) act like a giant lid,
trapping pollution close to the ground and enhancing regional coordination.

26
 CHAPTER 3. RESULTS

3.4.2 Step 2: Spatial Stability Analysis

Key Question: When winter amplifies pollution, do the hotspots migrate to different
countries, or do they stay in the exact same place?
To answer this, we performed a Spatial Shift Analysis, tracking how many cities
changed their LISA cluster classification between Summer and Winter.
The mathematical output revealed a Stability Index of 96.4%. This means 96.4%
of the cities retained their exact cluster classification across seasons. Out of 28 globally
distributed cities, only one city changed its status:

City Summer Cluster Winter Cluster Shift Type

Bangkok Not Significant High-High Emerged (NS → Cluster)

Table 3.4: Cities That Changed LISA Classification (Summer → Winter)

*Why Bangkok? Bangkok acts as a sensitive “buffer zone” on the edge of the Asian
pollution core. During summer, clean ocean monsoons push it out of the network. In
winter, shifting continental winds and regional transport pull it directly into the massive
High-High cluster.
Ultimately, air pollution has an extremely rigid core structure. The core hotspots—
Delhi, Beijing, Mumbai—remain

27
 CHAPTER 3. RESULTS

3.4.3 Step 3: Seasonal Variogram Analysis (The Heterogeneity

Paradox)
Key Question: If winter amplifies pollution, does the physical “Range” of a hotspot
expand to cover the whole world?
We fitted two separate Variogram models for Summer and Winter to compare their
spatial Ranges (the maximum distance of influence).

Figure 3.11: Seasonal Range Elasticity: The Winter curve reaches its Sill much faster
due to intense spatial heterogeneity.

The output revealed a fascinating paradox: The Winter Range actually contracts
by 13.6% (an absolute reduction of 905 km) compared to the Summer Range.
The mathematical parameters behind these curves provide the definitive proof of this
phenomenon:

Parameter Summer Winter Interpretation

Nugget 183.19 0 Micro-scale variance (noise)
Sill 689.41 2526.66 Maximum spatial variance
Range (km) 6653 5748 Distance to correlation cutoff
Sill/Range Ratio 0.104 0.440 Curve steepness (higher = faster plateau)
Expansion Factor 1.00 0.86× Seasonal range change

Table 3.5: Seasonal Variogram Parameters: Quantifying the Spatial Heterogeneity Para-
dox

28
 CHAPTER 3. RESULTS

*Note on Winter Nugget (0): The micro-scale variance drops to zero primarily due to
the preprocessing pipeline and the nature of the dataset. Seasonal aggregation effectively
smooths out random, short-term sensor noise. Additionally, given the sparse intercity
distance between the 28 globally distributed cities, hyperlocal fluctuations are structurally
filtered out.
Quantitative Breakdown:

• Range Contraction: The spatial range shrinks by 13.6% (an absolute reduction
of 905 km).

• Sill Explosion: The spatial variance (Sill) in winter is nearly 3.7 times higher
than in summer (2526.66 vs 689.41).

• Steepness Ratio: The Sill/Range ratio jumps from 0.104 to 0.440.

Urban air pollution operates as a geographically anchored network with seasonal elas-
ticity. The structural core is locked in place (96.4% stability), but winter tightens the
screws, amplifying the intensity and hyper-concentrating the danger into extreme regional
hotspots.
We now understand the spatial and temporal physics of pollution. However, the
atmosphere is driven by dozens of variables (Temperature, Humidity, Wind, PM2.5,
NO2). To feed this into an Early Warning System, how can we compress this chaotic,
multi-dimensional weather data into simple, predictable “Weather Regimes”?

29
 CHAPTER 3. RESULTS

3.5 Multivariate Regimes

To capture joint weather–pollution patterns, an 8-dimensional daily feature set is
constructed from standardized means of temperature, humidity, pressure, wind speed,
NO2 , SO2 , CO, and O3 for the core cities.

3.5.1 Multicollinearity Check (VIF)

Multicollinearity is assessed using variance inflation factors (VIF). All values are low
(VIF ≈ 1.26–1.84, highest for CO and NO2 ), indicating weak redundancy among pre-
dictors and supporting the use of the original variables for interpretation rather than
applying PCA (Table 3.6).

Table 3.6: VIF results for the 8 standardized drivers (core cities).

Variable VIF
co mean 1.84
no2 mean 1.72
so2 mean 1.46
wind mean 1.39
humid mean 1.38
press mean 1.38
o3 mean 1.26

3.5.2 Discover regimes with DBSCAN

DBSCAN is applied to identify typical regimes and isolate rare conditions. The pa-
rameters are set to minPts=5 and eps=1.5, guided by the elbow in the kNN-distance plot
(Figure 3.12). The curve remains relatively flat for most points and increases sharply near
the tail, suggesting one dominant dense baseline pattern and a small number of distant
observations (potential anomalies).

Figure 3.12: kNN distance plot (k = 5) used to select the DBSCAN eps.

30
 CHAPTER 3. RESULTS

The DBSCAN visualization (Figure 3.13) is consistent with this structure: one large
cluster dominates, with only a few isolated points. For interpretation, DBSCAN outputs
are mapped to two labels, Normal Baseline (main cluster) and Rare or Anomaly (noise
points). The figure is a 2D projection, while clustering is performed in the full 8D
standardized feature space.

Figure 3.13: DBSCAN clustering result for core cities (eps=1.5, minPts=5).

3.5.3 EWS evidence: do regimes change next-day extreme risk?

Regime context is evaluated by comparing the next-day extreme rate (extreme tomorrow)
across regimes. The Rare or Anomaly regime has a higher next-day extreme rate (0.286)
than Normal Baseline (0.189), corresponding to a gap of about +0.097 (Table 3.7).
This suggests that unusual multivariate conditions are associated with elevated next-
day risk. However, the rare regime includes only 42 city–days, so the result should be
interpreted cautiously and is formally tested in the predictive modeling section.

Table 3.7: Next-day extreme rate by DBSCAN regime (core cities, no PCA).

Weather Regime n Extreme rate Gap vs. baseline

Rare or Anomaly 42 0.286 0.097
Normal Baseline 1222 0.189 0.000

3.6 Predictive Validation

This section evaluates the Early Warning System (EWS) for predicting the binary tar-
get extreme tomorrow. Three logistic-regression specifications are compared on the same
time-aware train/test split: Model A (temporal baseline), Model B (A + Weather Regime),
and Model C (A + standardized continuous meteorology drivers).

31
 CHAPTER 3. RESULTS

3.6.1 Discrimination performance (ROC-AUC and PR-AUC)

Discrimination performance is reported using two threshold-free metrics: ROC-AUC
and PR-AUC (more informative under class imbalance).

ROC-AUC (ranking ability). Model C achieves the highest ROC-AUC (0.744),

indicating the best overall ranking of high-risk vs. low-risk days. The temporal baseline
Model A reaches ROC-AUC = 0.687. Adding the binary regime label does not improve
ranking: Model B is slightly below Model A (ROC-AUC = 0.680). Overall, the regime
label provides limited incremental discrimination once temporal memory features are
included, whereas the continuous 8 drivers provide a clear gain.

Table 3.8: Test ROC-AUC comparison for Models A–C.

Model ROC-AUC
A (Temporal baseline) 0.687
B (+ Weather Regime) 0.680
C (+ Raw 8 vars) 0.744

PR-AUC (performance under imbalance). The same ordering is observed with

PR-AUC. Model C provides the best precision–recall trade-off (PR-AUC = 0.557), while
Models A and B are similar (PR-AUC = 0.496 and 0.493, respectively). This suggests
that continuous meteorology/chemistry drivers help identify extreme-risk days more re-
liably when positives are relatively rare.

Table 3.9: Test PR-AUC comparison for Models A–C.

Model PR-AUC
A (Temporal baseline) 0.496
B (+ Weather Regime) 0.493
C (+ Raw 8 vars) 0.557

ROC curves (visual comparison). Figure 3.14 confirms the table results. The curve
for Model C lies above Models A and B over most of the range, meaning it achieves higher
sensitivity for the same false-positive rate. Models A and B nearly overlap, consistent
with their similar ROC-AUC values.

32
 CHAPTER 3. RESULTS

Figure 3.14: ROC curves on the test set for Models A–C. Model C dominates most of
the curve, consistent with its higher ROC-AUC.

3.6.2 Operational performance at a fixed threshold

To summarize practical warning behavior, confusion metrics are computed on the test
set at a fixed threshold τ = 0.45 (Table 3.10). Model C provides the best balance between
catching extreme events and limiting false alarms. Compared with Model A, Model C
increases true positives (TP: 46 vs 44) and reduces missed events (FN: 35 vs 37), while
keeping the same warning rate (0.357). Model B performs slightly worse than Model A
at the same threshold, consistent with its weaker ROC-AUC and PR-AUC.

Model TP FP TN FN Precision Recall F1 Warning Rate

A (Temporal) 44 46 125 37 0.489 0.543 0.515 0.357
B (+ Regime) 41 47 124 40 0.466 0.506 0.485 0.349
C (+ 8 vars) 46 44 127 35 0.511 0.568 0.538 0.357

Table 3.10: Confusion Metrics on Test Set (τ = 0.45)

Although Model C is the best performer, overall operational metrics remain moder-
ate (precision and recall around 0.5–0.57). This is expected because next-day extreme
events are relatively rare and are influenced by unobserved factors (e.g., abrupt emission
changes and local micro-meteorology), which limits perfect separation. In practice, the
threshold τ = 0.45 represents a trade-off: some false alarms (FP) are tolerated to reduce
missed extreme days (FN), which is often preferred in early warning settings. Over-
all, the strongest EWS performance is achieved by combining temporal memory features
with continuous meteorology/chemistry drivers (Model C). The regime label is useful for
interpretation but does not improve predictive accuracy in this evaluation setting.

33
Chapter 4

Conclusion and Future Work

4.1 Conclusion
This report developed an integrated framework to characterize spatio-temporal dy-
namics of urban air pollution and to support an Early Warning System (EWS) for pre-
dicting whether extreme tomorrow occurs. Statistical profiling indicates that daily AQI
is typically right-skewed and heavy-tailed, motivating a city-specific, data-driven extreme
definition based on the 80th percentile of daily AQI. Temporal analysis shows short-term
persistence, supporting operational memory features (Lag1, Delta AQI, extreme lag1).
Spatial and spatio-temporal modules highlight non-random geographic structure and sea-
sonal elasticity in clustering.
Multivariate regimes are discovered using DBSCAN on eight standardized meteorol-
ogy drivers. Rare/anomalous regime days exhibit higher next-day extreme risk than
baseline days, providing interpretable atmospheric context. Predictive validation shows
that the best performance is achieved by combining temporal memory with continu-
ous drivers (Model C), improving ROC-AUC and PR-AUC over the temporal baseline.
Adding the binary regime label alone does not improve discrimination, suggesting it is
more useful for interpretation than for prediction once temporal memory is included.

4.2 Limitations
The study period is limited (August 2025–January 2026), and evaluation uses a
single time-aware split rather than rolling backtesting. The percentile-based extreme
threshold may differ from health-based regulatory cutoffs. Predictive performance is also
constrained by unobserved drivers. DBSCAN regimes are summarized into baseline vs
rare/noise, which may oversimplify atmospheric diversity.

4.3 Future Work

Future work includes extending to multi-year data with rolling evaluation, testing
alternative extreme definitions, integrating additional predictors, and comparing more
flexible predictive models and regime-discovery methods while maintaining interpretabil-
ity.

34
References

[1] World Air Quality Index Project, “Aqicn: World air quality index project,” https:
//[Link]/api/, 2025.

[2] R. S. Bivand, E. Pebesma, and V. Gomez-Rubio, Applied Spatial Data Analysis with
R, 2nd ed. Springer, 2013.

[3] G. James, D. Witten, T. Hastie, and R. Tibshirani, An Introduction to Statistical

Learning: With Applications in R, 2nd ed. Springer, 2021.

[4] N. Cressie, Statistics for Spatial Data. John Wiley & Sons, 1993.

35