Hong Kong Pollution Data - 2025 Update

It’s been over 8 years since my original analysis of Hong Kong pollution data. Time for an update! Has air quality improved? What trends can we observe now with almost 27 years of data (1998-2025)?

In this post, we’ll:

1. Update the scraping tool to work with the current EPD website
2. Download fresh data for the Central/Western district
3. Analyze long-term trends and confirm that pollution is decreasing

Data downloaded from HK EPD.

Part 1: Updating the Scraper

The original Selenium-based scraper from 2018 no longer works - the EPD website has been completely redesigned and now uses Jakarta Server Faces (JSF). I’ve rewritten the scraper using Playwright, which handles the dynamic JavaScript forms better.

Key changes:

• Switched from Selenium to Playwright (headless mode, works over SSH)
• Handle JSF form submissions with hidden elements
• Download “By Station” format (wide format with all pollutants as columns)

# Install dependencies
# pip install playwright pandas numpy python-aqi
# playwright install chromium

# Download data for Central/Western station (1998-2025)
python hk_pollution_scraper.py \
    --stations central_western \
    --time-range hourly \
    --from 1998-01-01 \
    --to 2025-01-19 \
    --output central-western

The scraper automatically:

• Downloads yearly CSV files from EPD
• Converts units to EPA standard (CO to ppm, O3 to ppm, NO2/SO2 to ppb)
• Calculates 8-hour rolling averages for CO and O3
• Computes AQI using the python-aqi library
• Saves both hourly and daily AQI data

Please leave a comment if you need the script, i’ll upload it somewhere

Part 2: Loading the Data

import pandas as pd
import os

folder = 'central-western'
df = pd.read_csv(os.path.join(os.getcwd(), folder, 'aqi_hourly_full.csv'))
df.drop(df.columns[[0]], axis=1, inplace=True)
df.DATE = pd.to_datetime(df.DATE, format="%Y-%m-%d")
df.HOUR = pd.to_timedelta(df.HOUR, unit='h')
df['TIMESTAMP'] = df.DATE + df.HOUR
df.iloc[:,2:10] = df.iloc[:,2:10].apply(pd.to_numeric, errors='coerce')
dailyaqi = df.iloc[:,[0,9]].groupby(['DATE']).median().copy().T.squeeze()
hourlyaqi = df.loc[:,['TIMESTAMP','AQI']].groupby(['TIMESTAMP']).median().copy().T.squeeze()

Part 3: 27 Years of Hong Kong Air Quality

Let’s start with the full daily AQI time series from 1998 to 2025:

dailyaqi.plot(figsize=(16,6))

Several things stand out:

• The massive spike around 2010 reaching AQI ~500 was the infamous Asian dust storm from Mongolia/China
• There’s a clear regime change around 2011-2012 when PM2.5 monitoring began
• Recent years (2020-2025) show noticeably lower peaks

What’s Driving the AQI?

Let’s look at which pollutant is the primary driver of AQI each year:

aqi_origins = df.groupby(['DATE']).median().copy()
aqi_origins['AQI'] = aqi_origins.iloc[:,:6].max(numeric_only=True, axis=1)
aqi_origins['AQISource'] = aqi_origins.loc[:,['FSP','NO2','O3','RSP','SO2','CO_8h','O3_8h']].eq(
    aqi_origins['AQI'], axis=0).apply(
    lambda x: ','.join(aqi_origins.columns[0:7][x==x.max()]), axis=1)
aqi_origins.loc[aqi_origins['AQI'].isnull(),'AQISource'] = pd.np.nan
aqi_origins = aqi_origins.groupby(aqi_origins.index.year).apply(
    lambda x: x['AQISource'].value_counts(normalize=True))
mask = aqi_origins > 0.025
aqi_origins = aqi_origins.loc[mask]
aqi_origins = aqi_origins.unstack().fillna(0)
aqi_origins.plot.bar(stacked=True, figsize=(16,8))

The story is clear:

• 1998-2010: RSP (PM10) was the main driver, with some NO2 contribution
• 2011: Transition year - mix of RSP, NO2, and the new FSP (PM2.5) measurement
• 2012-2025: FSP (PM2.5) completely dominates as the primary pollutant

This transition coincides with Hong Kong starting to measure PM2.5 (Fine Suspended Particulates) around 2011-2012. PM2.5 wasn’t being measured before, so it couldn’t be the AQI driver even if it was present in the air.

Year-by-Year Distribution

Box plots show the distribution of daily AQI values per year:

groups = dailyaqi.groupby(dailyaqi.index.year)
groupstack = pd.concat([pd.DataFrame(x[1].values) for x in groups], axis=1)
groupstack.columns = [str(x) for x in groups.groups.keys()]
groupstack.boxplot(figsize=(16,6))

The 2010 outlier is clearly visible. More importantly, the median AQI (green line) and the interquartile range both appear to be declining in recent years.

The Last Decade: 2013-2025

Let’s zoom in on recent years:

dailyaqi['2013':].plot(style='k.', figsize=(16,6))

The downward trend becomes more apparent. The ceiling of daily values has clearly dropped - we’re seeing fewer days above AQI 150, and many more days in the “Good” range (below 50).

Seasonality: Heatmap View

A heatmap of the last 8 years (2018-2025) by day of year:

Winter months (left side, days 0-90 and right side, days 300-365) consistently show higher pollution (yellow/green), while summer months (days 150-250, roughly June-September) are cleaner (dark purple).

Weekly Seasonality

weeks = dailyaqi['2018':].groupby(dailyaqi['2018':].index.isocalendar().week)
weekstack = pd.concat([pd.DataFrame(x[1].values) for x in weeks], axis=1)
weekstack.columns = [str(x) for x in weeks.groups.keys()]
weekstack.boxplot(figsize=(16,6))

Weeks 25-35 (late June through early September) have the lowest pollution - this is the summer monsoon season when southerly winds bring cleaner maritime air. Winter weeks (1-10 and 45-52) see higher pollution from northerly winds carrying continental pollution.

Monthly Pattern

months = dailyaqi['2018':].groupby(dailyaqi['2018':].index.month)
monthstack = pd.concat([pd.DataFrame(x[1].values) for x in months], axis=1)
monthstack.columns = [str(x) for x in months.groups.keys()]
monthstack.boxplot(figsize=(12,6))

The pattern is even clearer by month:

• January-March: Higher pollution (median ~70-75)
• June-September: Cleanest months (median ~25-30)
• October-December: Rising again

This aligns with Hong Kong’s monsoon patterns - summer southwest monsoons bring clean oceanic air, while winter northeast monsoons bring pollution from mainland China.

Part 4: Additional Analysis

Long-Term Trend: Average AQI by Year

yearly_avg = dailyaqi.groupby(dailyaqi.index.year).mean()
yearly_avg.plot(kind='bar', figsize=(14,4), title='Average AQI by Year', color='steelblue')
plt.axhline(y=yearly_avg.mean(), color='red', linestyle='--', label=f'Overall mean: {yearly_avg.mean():.1f}')

The bar chart clearly shows:

• 1998-2010: Consistently low AQI (35-50 range) - but this is misleading as PM2.5 wasn’t measured
• 2011-2015: Spike to 60-90 after PM2.5 monitoring began
• 2016-2025: Gradual decline back toward 50

Rolling Averages: The True Trend

dailyaqi.rolling(365).mean().plot(label='365-day rolling mean', linewidth=2)
dailyaqi.rolling(30).mean().plot(alpha=0.3, label='30-day rolling mean')

The 365-day rolling average (blue line) tells the complete story:

• Stable around 45 from 1998-2010
• Sharp rise to peak of ~90 around 2013 (after PM2.5 monitoring)
• Steady decline since 2014, now back to ~50

Hour of Day Pattern

hourly_pattern = df.groupby('HOUR')['AQI'].median()
hourly_pattern.plot(kind='bar', figsize=(12,4), title='Median AQI by Hour of Day')

Interestingly, there’s not a dramatic rush-hour spike at Central/Western (a general station, not roadside). The pattern shows:

• Lowest: Early morning (3-7 AM), around AQI 44-45
• Highest: Afternoon/evening (3-9 PM), around AQI 54-55
• Difference of only ~10 AQI points

This suggests the pollution at this station is more influenced by regional/background levels than local traffic.

Day of Week Pattern

df['DOW'] = df['TIMESTAMP'].dt.dayofweek
dow_aqi = df.groupby('DOW')['AQI'].median()

Surprisingly, there’s almost no weekday/weekend difference:

• Weekday average: 51
• Weekend average: 51.5

This reinforces that the Central/Western station measures regional pollution rather than local traffic emissions. Roadside stations like Causeway Bay would show a bigger difference.

COVID-19 Impact: 2019 vs 2020 vs 2021

comparison = pd.DataFrame({
    '2019': monthly_2019.values,
    '2020': monthly_2020.values,
    '2021': monthly_2021.values
}, index=['Jan','Feb','Mar','Apr','May','Jun','Jul','Aug','Sep','Oct','Nov','Dec'])
comparison.plot(kind='bar')

COVID-19 impact is visible:

• 2019 median AQI: 66.0
• 2020 median AQI: 59.0
• 2021 median AQI: 56.0

The reduction during 2020-2021 lockdowns is clear, especially in winter months when the difference is most pronounced. Summer months show less difference as they’re already clean due to monsoon patterns.

Exceedance Days: How Often is Air “Unhealthy”?

thresholds = {
    'Unhealthy for Sensitive (>100)': 100,
    'Unhealthy (>150)': 150,
    'Very Unhealthy (>200)': 200
}

This is perhaps the most striking chart:

• 2013 peak: 137 days above AQI 100, 45+ days above 150
• 2020-2025: Only 5-15 days above AQI 100, almost zero above 150

The number of “bad air days” has dropped by 90% since the worst years.

Seasonal Decomposition

Breaking down the time series into trend and seasonal components:

The three panels show:

1. Top: Raw monthly median AQI - the noisy signal
2. Middle: Trend component (12-month rolling average) - clearly shows the rise and fall
3. Bottom: Seasonal component - consistent winter peaks and summer troughs throughout the entire period

Era Comparison

Grouping into distinct periods:

decades = {
    '1998-2004': dailyaqi['1998':'2004'],
    '2005-2011': dailyaqi['2005':'2011'],
    '2012-2018': dailyaqi['2012':'2018'],
    '2019-2025': dailyaqi['2019':'2025']
}

The 2012-2018 period (post-PM2.5 monitoring) was the worst, but 2019-2025 shows significant improvement - AQI has dropped 30% from the 2012-2018 peak.

The Key Finding: Statistically Significant Improvement

Focusing on 2016-2025 (after PM2.5 monitoring stabilized):

from scipy import stats
yearly_avg_recent = yearly_avg[yearly_avg.index >= 2016]
slope, intercept, r_value, p_value, std_err = stats.linregress(years, yearly_avg_recent.values)

Statistical Results:

• Trend: -2.82 AQI per year
• R-squared: 0.89 (very strong fit)
• P-value: 0.000 (highly significant)

Period Comparison:

• 2016-2020 average: 64.4 AQI
• 2021-2025 average: 49.7 AQI
• Improvement: 14.7 points (23% reduction)

Conclusions

After 27 years of data and comprehensive analysis:

1. PM2.5 dominates since 2012 - Fine Suspended Particulates drive nearly 100% of high AQI days

2. Strong, consistent seasonality - Summer (June-September) is ~60% cleaner than winter due to monsoon patterns

3. No significant weekday/weekend or rush-hour effect - Central/Western measures regional pollution, not local traffic

4. COVID-19 had measurable impact - 10-15% reduction in AQI during 2020-2021

5. 90% fewer “bad air days” - Days exceeding AQI 100 dropped from 137 (2013) to under 15 (recent years)

6. Statistically significant improvement since 2016 - AQI declining by 2.8 points per year, with 23% overall reduction

The improvement is likely due to:

• Stricter emissions controls in Hong Kong
• Reduced industrial activity in the Pearl River Delta
• Cleaner vehicle fleet (more EVs, tighter emission standards)
• Some effect from reduced activity during COVID-19

Hong Kong’s air quality is genuinely improving. The data shows real, measurable, statistically significant progress over the last decade. While PM2.5 remains the primary concern, the trend is encouraging.

Code and data available on GitHub. Updated scraper works with the 2025 EPD website.