11.2. Exploring and Cleaning AQS Sensor Data¶

Let’s briefly take stock of where we are in the analysis. Our plan is roughly to:

Find a list of possibly collocated AQS and PurpleAir sensors.
Contact AQS sites to find truly collocated sensor pairs.
Explore and clean AQS measurements for one sensor.
Explore and clean PurpleAir measurements for the other sensor in the pair, then join the measurements together.
Join all the measurements together for all the sensor pairs.
Fit a model to make PurpleAir measurements match AQS measurements.

We finished step 1 in the previous section (Section 11.1). We skipped step 2; we decided to reuse Barkjohn’s list of truly collocated sensor pairs rather than contacting the AQS sites ourselves.

In this section, we’ll proceed to step 3 of the analysis plan: to explore and clean data from one AQS sensor.

We picked a sensor pair from Barkjohn’s list of collocated sensors. The pair is located in Sacramento, California. The AQS sensor ID is 06-067-0010, and the PurpleAir sensor name is AMTS_TESTINGA.

The AQS provides a website and API to download sensor data 1. For this sensor, Barkjohn’s analysis used data from May 20, 2018 to Dec 29, 2019. We downloaded the daily measurements for these dates into the data/aqs_06-067-0010.csv file. We’ll begin by loading this file into pandas.

aqs = pd.read_csv('data/aqs_06-067-0010.csv')
aqs.head(2)

	state_code	county_code	site_number	parameter_code	...	city	cbsa_code	cbsa	date_of_last_change
0	6	67	10	88101	...	Sacramento	40900	Sacramento--Roseville--Arden-Arcade, CA	2021-09-24
1	6	67	10	88101	...	Sacramento	40900	Sacramento--Roseville--Arden-Arcade, CA	2021-09-24

2 rows × 31 columns

As usual, this dataframe has too many columns to display in the notebook. We’ll display the first row of data to see all the columns.

display_df(aqs.iloc[0].to_frame(), rows=31)

	0
state_code	6
county_code	67
site_number	10
parameter_code	88101
poc	1
latitude	38.57
longitude	-121.49
datum	NAD83
parameter	PM2.5 - Local Conditions
sample_duration	24 HOUR
pollutant_standard	PM25 24-hour 2006
date_local	2018-12-31
units_of_measure	Micrograms/cubic meter (LC)
event_type	None
observation_count	1
observation_percent	100.0
validity_indicator	Y
arithmetic_mean	3.1
first_max_value	3.1
first_max_hour	0
aqi	13
method_code	145
method	R & P Model 2025 PM-2.5 Sequential Air Sampler...
local_site_name	Sacramento-1309 T Street
site_address	1309 T ST., SACRAMENTO, CA. 95814
state	California
county	Sacramento
city	Sacramento
cbsa_code	40900
cbsa	Sacramento--Roseville--Arden-Arcade, CA
date_of_last_change	2021-09-24

From the data dictionary for this data 2, we can find out that the arithmetic_mean column corresponds to the actual PM2.5 measurements. Some AQS sensors take a measurement every hour. For our analysis, we downloaded the 24-hour averages (the arithmetic mean) of the hourly sensor measurements.

Before we can use this data for analysis, we’ll perform validity checks with a combination of EDA and visualization, then clean the data where necessary.

We will:

Check and correct the granularity of the data.
Remove unneeded columns.
Check validity of the date_local column.
Check validity of the arithmetic_mean column.

There are many more checks we could perform on this data. If we were planning to publish this work more broadly, we would be as thorough as possible. For the sake of brevity, we’ve chosen a few important validity checks that specifically reinforce ideas we’ve covered in data wrangling, EDA, and visualization.

11.2.1. What’s the Granularity?¶

We would like each row of our data to correspond to a single date with an average PM2.5 reading for that date. Is this the case for the aqs dataframe? One simple way to check is to see whether there are repeat values in the date_local column.

aqs['date_local'].value_counts()

2019-05-12    12
2019-11-14    12
2019-09-03    12
              ..
2019-01-30    12
2019-12-23    12
2019-05-27    12
Name: date_local, Length: 189, dtype: int64

Indeed, there are 12 rows for each date in aqs. So, the granularity is not at the individual date level. To figure out why this happens, we can filter the table to a single date, then look for columns that have different values within that date. If we do this, we find that the pollutant_standard and event_type columns differ within a date. Here, we display these columns for the date 2018-12-31:

one_date = (aqs.query('date_local == "2018-12-31"')
 [['date_local', 'pollutant_standard', 'event_type', 'arithmetic_mean']]
)
display_df(one_date, rows=12)

	date_local	pollutant_standard	event_type	arithmetic_mean
0	2018-12-31	PM25 24-hour 2006	None	3.1
1	2018-12-31	PM25 24-hour 2006	Included	3.1
140	2018-12-31	PM25 Annual 2006	None	3.1
141	2018-12-31	PM25 Annual 2006	Included	3.1
280	2018-12-31	PM25 24-hour 2012	None	3.1
281	2018-12-31	PM25 24-hour 2012	Included	3.1
420	2018-12-31	PM25 Annual 2012	None	3.1
421	2018-12-31	PM25 Annual 2012	Included	3.1
560	2018-12-31	PM25 24-hour 1997	None	3.1
561	2018-12-31	PM25 24-hour 1997	Included	3.1
700	2018-12-31	PM25 Annual 1997	None	3.1
701	2018-12-31	PM25 Annual 1997	Included	3.1

From the data dictionary, we learn that there are multiple standards for computing the final measurements from the raw sensor data. The pollutant_standard column contains the name of each standard. The event_type column marks whether data measured during “exceptional events” are included in the measurement.

However, for our data these columns seem to have no effect on the PM2.5 measurements. For example, in the 2018-12-31 data above the PM2.5 measurements are all identical. We can verify this for the rest of the data by checking that the maximum PM2.5 minus the minimum PM2.5 for each date is equal to zero.

(aqs
 .groupby('date_local')
 ['arithmetic_mean']
 .agg(np.ptp) # np.ptp computes max() - min()
 .value_counts()
)
# For all 189 dates, the max PM2.5 - min PM2.5 is 0

0.0    189
Name: arithmetic_mean, dtype: int64

So, we can simply take the first PM2.5 measurement for each date.

def rollup_dates(df):
    return (
        df.groupby('date_local')
        .first()
        .reset_index()
    )

aqs = (pd.read_csv('data/aqs_06-067-0010.csv')
       .pipe(rollup_dates))
aqs.head(2)

	date_local	state_code	county_code	site_number	...	city	cbsa_code	cbsa	date_of_last_change
0	2018-05-20	6	67	10	...	Sacramento	40900	Sacramento--Roseville--Arden-Arcade, CA	2021-09-24
1	2018-05-23	6	67	10	...	Sacramento	40900	Sacramento--Roseville--Arden-Arcade, CA	2021-09-24

2 rows × 31 columns

This data cleaning step gives us the desired granularity of aqs: every row in aqs represents a single date, with an average PM2.5 measurement for that date.

11.2.2. Removing Unneeded Columns¶

We plan to match the PM2.5 measurements in the aqs dataframe with the PurpleAir PM2.5 measurements for each date. To simplify the data, we’ll subset out the date and PM2.5 columns and rename the PM2.5 column so that it’s easier to understand.

def subset_cols(df):
    subset = df[['date_local', 'arithmetic_mean']]
    return subset.rename(columns={'arithmetic_mean': 'pm25'})

aqs = (pd.read_csv('data/aqs_06-067-0010.csv')
       .pipe(rollup_dates)
       .pipe(subset_cols))
aqs

	date_local	pm25
0	2018-05-20	6.5
1	2018-05-23	2.3
2	2018-05-29	11.8
...	...	...
186	2019-12-23	5.7
187	2019-12-26	2.0
188	2019-12-29	19.5

189 rows × 2 columns

11.2.3. Checking the Validity of `date_local`¶

Let’s take a closer look at the date_local column. We can already see that there are gaps in dates where there are no PM2.5 readings.

# The table is sorted by `date_local`, so we see that there are missing dates
# between 2018-05-20 and 2018-05-23, for example.
aqs

	date_local	pm25
0	2018-05-20	6.5
1	2018-05-23	2.3
2	2018-05-29	11.8
...	...	...
186	2019-12-23	5.7
187	2019-12-26	2.0
188	2019-12-29	19.5

189 rows × 2 columns

We can check that the date_local column contains strings:

# Python strings are recorded as the `object` type in pandas
aqs.dtypes

date_local     object
pm25          float64
dtype: object

aqs['date_local'].iloc[0]

'2018-05-20'

However, we’d like to know: are all the strings valid dates? If so, we can parse the dates as pd.Timestamp objects which will make it easier to compute the dates that are missing from the data.

The date_local values appear to come in the format: YYYY-MM-DD, where YYYY, MM, and DD correspond to the year, month, and day. The pd.to_datetime() method can parse the date strings into objects, and we can pass in the format of the dates as a date format string 3.

# This is the Python representation of the YYYY-MM-DD format
date_format = '%Y-%m-%d'

pd.to_datetime(aqs['date_local'], format=date_format)

   2018-05-20
   2018-05-23
   2018-05-29
         ...    
 2019-12-23
 2019-12-26
 2019-12-29
Name: date_local, Length: 189, dtype: datetime64[ns]

We see that the dtype of this series is datetime64[ns] rather than object, which means that the values are pd.TimeStamp objects rather than strings. Also, the method runs without erroring, indicating that all the strings matched the format.

Now, we can reassign the date_local column as pd.TimeStamps.

def parse_dates(df):
    date_format = '%Y-%m-%d'
    timestamps = pd.to_datetime(df['date_local'], format=date_format)
    return df.assign(date_local=timestamps)

aqs = (pd.read_csv('data/aqs_06-067-0010.csv')
       .pipe(rollup_dates)
       .pipe(subset_cols)
       .pipe(parse_dates))
aqs

	date_local	pm25
0	2018-05-20	6.5
1	2018-05-23	2.3
2	2018-05-29	11.8
...	...	...
186	2019-12-23	5.7
187	2019-12-26	2.0
188	2019-12-29	19.5

189 rows × 2 columns

aqs.dtypes

date_local    datetime64[ns]
pm25                 float64
dtype: object

Note

Just because the dates can be parsed doesn’t mean that the dates are immediately ready to use for further analysis. For instance, the string 9999-01-31 can be parsed into a pd.TimeStamp, but the date isn’t valid.

Now that the date_local contains timestamps, we can calculate how many dates are missing. We’ll find the number of days between the earliest and latest date—this corresponds to the maximum number of measurements we could have recorded.

date_range = aqs['date_local'].max() - aqs['date_local'].min()
date_range

Timedelta('588 days 00:00:00')

# Subtracting timestamps give Timedelta objects, which have a few useful
# properties like:
date_range.days

print(f'We have {len(aqs)} / {date_range.days} measurements, '
      f'or {len(aqs) / date_range.days:.0%} of the dates possible.')

We have 189 / 588 measurements, or 32% of the dates possible.

There are many dates missing from the data. However, when we combine this data with the data from the other AQS and PurpleAir sensors, we expect that we’ll have enough data to fit a model.

There are other modeling scenarios where this amount of missing data might cause more concern, and we’ve included an exercise at the end of this chapter to think about potential issues.

11.2.4. Checking the Validity of `pm25`¶

There are a few checks we can perform on the PM2.5 measurements. First, PM2.5 can’t go below 0. Second, we can look for abnormally high PM2.5 values and see whether they correspond to major events like a wildfire.

One simple way to perform these checks is to plot the PM2.5 against the date.

sns.lineplot(data=aqs, x='date_local', y='pm25')
# Rotate x-axis labels to avoid overlap
plt.xticks(rotation=45);

We can see that the PM2.5 measurements don’t go below 0.

We also a large spike in PM2.5 around November of 2018. This sensor is located in Sacramento, CA—was there a fire around that area?

Indeed, November 8, 2018 marks the start of the Camp Fire, the “deadliest and most destructive wildfire in California history” [Bureau, 2018]. The fire started just 80 miles north of Sacramento, so this AQS sensor captured the dramatic spike in PM2.5.

11.2.5. Next Steps¶

We’ve cleaned and explored the data for an AQS sensor. In the next section, we’ll do the same for its collocated PurpleAir sensor.

1: https://aqs.epa.gov/aqsweb/documents/data_api.html
2: https://aqs.epa.gov/aqsweb/airdata/FileFormats.html#_hourly_data_files
3: For more about date format strings, see the quick reference at https://strftime.org/ or the Python documentation.

Principles and Techniques of Data Science

Exploring and Cleaning AQS Sensor Data

Contents

11.2. Exploring and Cleaning AQS Sensor Data¶

11.2.1. What’s the Granularity?¶

11.2.2. Removing Unneeded Columns¶

11.2.3. Checking the Validity of `date_local`¶

11.2.4. Checking the Validity of `pm25`¶

11.2.5. Next Steps¶

Principles and Techniques of Data Science

Exploring and Cleaning AQS Sensor Data

Contents

11.2. Exploring and Cleaning AQS Sensor Data¶

11.2.1. What’s the Granularity?¶

11.2.2. Removing Unneeded Columns¶

11.2.3. Checking the Validity of date_local¶

11.2.4. Checking the Validity of pm25¶

11.2.5. Next Steps¶

11.2.3. Checking the Validity of `date_local`¶

11.2.4. Checking the Validity of `pm25`¶