Data wrangling, grouping and aggregation#

Next, we will continue working with weather data, but expand our analysis to cover longer periods of data from Finland. In the following, you will learn various useful techniques in pandas to manipulate, group and aggregate the data in different ways that are useful when extracting insights from your data. In the end, you will learn how to create an automated data analysis workflow that can be repeated with multiple input files having a similar structure. As a case study, we will investigate whether January 2020 was the warmest month on record also in Finland, as the month was the warmest one on record globally [1].

Cleaning data while reading#

In this section we are using weather observation data from Finland that was downloaded from NOAA (see Datasets chapter for further details). The input data is separated with varying number of spaces (i.e., fixed width). The first lines and columns of the data look like following:

  USAF  WBAN YR--MODAHRMN DIR SPD GUS CLG SKC L M H  VSB MW MW MW MW AW  ...
029440 99999 190601010600 090   7 *** *** OVC * * *  0.0 ** ** ** ** **  ...
029440 99999 190601011300 ***   0 *** *** OVC * * *  0.0 ** ** ** ** **  ...
029440 99999 190601012000 ***   0 *** *** OVC * * *  0.0 ** ** ** ** **  ...
029440 99999 190601020600 ***   0 *** *** CLR * * *  0.0 ** ** ** ** **  ...

By looking at the data, we can notice a few things that we need to consider when reading the data:

  1. Delimiter: The columns are separated with a varying amount of spaces which requires using some special tricks when reading the data with pandas read_csv() function

  2. NoData values: NaN values in the NOAA data are coded with varying number of * characters, hence, we need to be able to instruct pandas to interpret those as NaNs.

  3. Many columns: The input data contains many columns (altogether 33). Many of those do not contain any meaningful data for our needs. Hence, we should probably ignore the unnecessary columns already at this stage.

Handling and cleaning heterogeneous input data (such as our example here) can be done after reading in the data. However, in many cases, it is actually useful to do some cleaning and preprocessing already when reading the data. In fact, that is often much easier to do. In our case, we can read the data with varying number of spaces between the columns (1) by using a parameter delim_whitespace=True (alternatively, specifying sep='\s+' would work). For handling the NoData values (2), we can tell pandas to consider the * characters as NaNs by using a paramater na_values and specifying a list of characters that should be converted to NaNs. Hence, in this case we can specify na_values=['*', '**', '***', '****', '*****', '******'] which will then convert the varying number of * characters into NaN values. Finally, we can limit the number of columns that we read (3) by using the usecols parameter, which we already used previously. In our case, we are interested in columns that might be somehow useful to our analysis, including the station name, timestamp, and data about temperatures: 'USAF', 'YR--MODAHRMN', 'TEMP', 'MAX', 'MIN'. Achieving all these things is pretty straightforward using the read_csv() function:

import pandas as pd

# Define relative path to the file
fp = "data/029820.txt"

# Read data using varying amount of spaces as separator,
# specifying '*' characters as NoData values,
# and selecting only specific columns from the data
data = pd.read_csv(
    fp,
    delim_whitespace=True,
    na_values=["*", "**", "***", "****", "*****", "******"],
    usecols=["USAF", "YR--MODAHRMN", "TEMP", "MAX", "MIN"],
)
/var/folders/f7/rhmqxfmx40s4yv9bhh7skq4m0000gp/T/ipykernel_67646/2184805043.py:9: FutureWarning: The 'delim_whitespace' keyword in pd.read_csv is deprecated and will be removed in a future version. Use ``sep='\s+'`` instead
  data = pd.read_csv(

Let’s see now how the data looks by printing the first five rows with the head() function:

data.head()
USAF YR--MODAHRMN TEMP MAX MIN
0 29820 190601010600 34.0 NaN NaN
1 29820 190601011300 32.0 NaN NaN
2 29820 190601012000 30.0 NaN NaN
3 29820 190601020600 33.0 NaN NaN
4 29820 190601021300 35.0 NaN NaN

Perfect, looks good. We have skipped a bunch of unnecessary columns and also the asterisk (*) characters have been correctly converted to NaN values.

Renaming columns#

Let’s take a closer look at the column names of our DataFrame:

print(data.columns)
Index(['USAF', 'YR--MODAHRMN', 'TEMP', 'MAX', 'MIN'], dtype='object')

As we see, some of the column names are a bit awkward and difficult to interpret (a description for the columns is available in the metadata data/3505doc.txt). Luckily, it is easy to alter labels in a pandas DataFrame using the rename() function. In order to change the column names, we need to tell pandas how we want to rename the columns using a dictionary that converts the old names to new ones. As you probably remember from Chapter 1, a dictionary is a specific data structure in Python for storing key-value pairs. We can define the new column names using a dictionary where we list “key: value” pairs in following manner:

  • USAF: STATION_ID

  • YR--MODAHRMN: TIME

  • TEMP: TEMP_F

Hence, the original column name (e.g. YR--MODAHRMN) is the dictionary key which will be converted to a new column name TIME (which is the value). The temperature values in our data file is again represented in Fahrenheit. We will soon convert these temperatures to Celsius. Hence, in order to avoid confusion with the columns, let’s rename the column TEMP to TEMP_F. Also the station number USAF is much more intuitive if we call it STATION_ID. Let’s create a dictionary for the new column names:

new_names = {
    "USAF": "STATION_ID",
    "YR--MODAHRMN": "TIME",
    "TEMP": "TEMP_F",
}
new_names
{'USAF': 'STATION_ID', 'YR--MODAHRMN': 'TIME', 'TEMP': 'TEMP_F'}

Our dictionary looks correct, so now we can change the column names by passing that dictionary using the parameter columns in the rename() function:

data = data.rename(columns=new_names)
data.columns
Index(['STATION_ID', 'TIME', 'TEMP_F', 'MAX', 'MIN'], dtype='object')

Perfect, now our column names are easier to understand and use.

Using functions with pandas#

Now it’s time to convert those temperatures from Fahrenheit to Celsius. We have done this many times before, but this time we will learn how to apply our own functions to data in a pandas DataFrame. We will define a function for the temperature conversion, and apply this function for each Celsius value on each row of the DataFrame. Output celsius values should be stored in a new column called TEMP_C. But first, it is a good idea to check some basic properties of our new input data before proceeding with data analysis:

# First rows
data.head(2)
STATION_ID TIME TEMP_F MAX MIN
0 29820 190601010600 34.0 NaN NaN
1 29820 190601011300 32.0 NaN NaN
# Last rows
data.tail(2)
STATION_ID TIME TEMP_F MAX MIN
198332 29820 201910012200 47.0 NaN NaN
198333 29820 201910012300 46.0 NaN NaN
# Data types
data.info()
<class 'pandas.core.frame.DataFrame'>
RangeIndex: 198334 entries, 0 to 198333
Data columns (total 5 columns):
 #   Column      Non-Null Count   Dtype  
---  ------      --------------   -----  
 0   STATION_ID  198334 non-null  int64  
 1   TIME        198334 non-null  int64  
 2   TEMP_F      197916 non-null  float64
 3   MAX         29868 non-null   float64
 4   MIN         29536 non-null   float64
dtypes: float64(3), int64(2)
memory usage: 7.6 MB

Nothing suspicous for the first and last rows, but here with info() we can see that the number of observations per column seem to be varying if you compare the Non-Null Count information to the number of entries in the data (N=198334). Only station number and time seem to have data on each row. All other columns seem to have some missing values. This is not necessarily anything dangerous, but good to keep in mind. Let’s still look at the descriptive statistics:

# Descriptive stats
data.describe()
STATION_ID TIME TEMP_F MAX MIN
count 198334.0 1.983340e+05 197916.000000 29868.000000 29536.000000
mean 29820.0 1.990974e+11 43.717845 46.405852 40.537446
std 0.0 2.691914e+09 14.306138 14.446547 14.350235
min 29820.0 1.906010e+11 -28.000000 -20.000000 -28.000000
25% 29820.0 1.979011e+11 34.000000 36.000000 31.000000
50% 29820.0 1.997061e+11 43.000000 45.000000 40.000000
75% 29820.0 2.013113e+11 55.000000 59.000000 52.000000
max 29820.0 2.019100e+11 90.000000 102.000000 77.000000

By looking at the TEMP_F values (Fahrenheit temperatures), we can confirm that our measurements seems more or less valid because the value range of the temperatures makes sense, i.e. there are no outliers such as extremely high MAX values or low MIN values. It is always a good practice to critically check your data before doing any analysis, as it is possible that your data may include incorrect values, e.g. due to a sensor malfunction or human error.

Defining a function#

Now we are sure that our data looks okay, and we can start our temperature conversion process by first defining our temperature conversion function from Fahrenheit to Celsius. Pandas can use regular functions, hence you can define functions for pandas exactly in the same way as you would do normally (as we learned in Chapter 1). Hence, let’s define a function that converts Fahrenheits to Celsius:

def fahr_to_celsius(temp_fahrenheit):
    """Function to convert Fahrenheit temperature into Celsius.

    Parameters
    ----------

    temp_fahrenheit: int | float
        Input temperature in Fahrenheit (should be a number)

    Returns
    -------

    Temperature in Celsius (float)
    """

    # Convert the Fahrenheit into Celsius
    converted_temp = (temp_fahrenheit - 32) / 1.8

    return converted_temp

Now we have the function defined and stored in memory. At this point it is good to test the function with some known value:

fahr_to_celsius(32)
0.0

32 Fahrenheits is indeed 0 Celsius, so our function seem to be working correctly.

Using a function by iterating over rows#

Next we will learn how to use our function with data stored in a pandas DataFrame. We will first apply the function row-by-row using a for loop and then we will learn a more efficient way of applying the function to all rows at once.

Looping over rows in a DataFrame can be done in a couple of different ways. A common approach is to use the iterrows() method which loops over the rows as index-Series pairs. In other words, we can use the iterrows() method together with a for loop to repeat a process for each row in a Pandas DataFrame. Please note that iterating over rows this way is a rather inefficient approach, but it is still useful to understand the logic behind how this works. When using the iterrows() method it is important to understand that iterrows() accesses not only the values of one row, but also the index of the row. Let’s start with a simple example for loop that goes through each row in our DataFrame.

# Iterate over the rows
for idx, row in data.iterrows():
    # Print the index value
    print("Index:", idx)

    # Print the temperature from the row
    print("Temp F:", row["TEMP_F"], "\n")

    break
Index: 0
Temp F: 34.0 

We can see that the idx variable indeed contains the index value at position 0 (the first row) and the row variable contains all the data from that given row stored as a pandas Series. Also, notice that when developing a for loop you do not always need to iterate through the entire loop if you just want to test things out. Using the break statement in Python terminates a loop whenever it is placed inside the loop. Here we used it to check out the values on the first row of the DataFrame. This allows us to test the code logic without printing thousands of values to the screen!

Next, let’s create an empty column TEMP_C for the Celsius temperatures and update the values in that column using the fahr_to_celsius() function that we defined earlier. For updating the value in the DataFrame, we can use the at method that we already used earlier in this chapter. This time, however, we will use the itertuples() method to access the rows in the DataFrame. The itertuples() method works similarly to iterrows(), except it returns only the row values without the index. In addition, the returned values are not a pandas Series, but instead itertuples() returns a named tuple data type. As a result, when using itertuples() accessing the row values needs to be done a bit differently. Remember, a tuple is like a list but immutable and a “named tuple” is a special kind of tuple object that adds the ability to access the values by name instead of position index. Hence, we can access the TEMP_F value in a given row using row.TEMP_F (in contrast to how we accessed the value in the previous code above). We will not work with named tuples in the rest of the book, but more information can be found in the Python documentation for named tuples [2].

Let’s see an example of how to use the itertuples() method.

# Create an empty column for the output values
data["TEMP_C"] = 0.0

# Iterate over the rows
for row in data.itertuples():
    # Convert the Fahrenheit to Celsius
    # Notice how we access the row value
    celsius = fahr_to_celsius(row.TEMP_F)

    # Update the value for 'Celsius' column with the converted value
    # Notice how we can access the Index value
    data.at[row.Index, "TEMP_C"] = celsius
# Check the result
data.head()
STATION_ID TIME TEMP_F MAX MIN TEMP_C
0 29820 190601010600 34.0 NaN NaN 1.111111
1 29820 190601011300 32.0 NaN NaN 0.000000
2 29820 190601012000 30.0 NaN NaN -1.111111
3 29820 190601020600 33.0 NaN NaN 0.555556
4 29820 190601021300 35.0 NaN NaN 1.666667
# What does our row look like?
row._asdict()
{'Index': 198333,
 'STATION_ID': 29820,
 'TIME': 201910012300,
 'TEMP_F': 46.0,
 'MAX': nan,
 'MIN': nan,
 'TEMP_C': 0.0}

Okay, now we have iterated over our data and updated the temperatures in Celsius to TEMP_C column by using our fahr_to_celsius() function. The values look correct as 32 degrees Fahrenheit indeed is 0 Celsius degrees, as can be seen on the second row. We also have the last row of our DataFrame in the code above, which is a named tuple that has been converted to the more familiar dictionary data type using the _asdict() method for named tuples.

Before moving to other more efficient ways to use functions with pandas DataFrames, we should note a few things about the approaches above. We demonstrated use of the itertuples() method for looping over the values because it is significantly faster than iterrows() (can be around 100x faster). We also used .at to assign the value to the DataFrame because it is designed to access single values more efficiently than .loc, which can access also groups of rows and columns. That said, you could have also simply used data.loc[idx, new_column] = celsius to achieve the same result as both examples above. It is just slower.

Using a function with apply#

Although using for loop with itertuples() can be fairly efficient, pandas DataFrames and Series have a dedicated method called apply() for applying functions on columns (or rows). apply() is typically faster than itertuples(), especially if you have large number of rows, such as in our case. When using apply(), we pass the function that we want to use as an argument. Let’s start by applying the function to the TEMP_F column that contains the temperature values in Fahrenheit:

data["TEMP_F"].apply(fahr_to_celsius)
0         1.111111
1         0.000000
2        -1.111111
3         0.555556
4         1.666667
            ...   
198329    8.333333
198330    8.333333
198331    8.333333
198332    8.333333
198333    7.777778
Name: TEMP_F, Length: 198334, dtype: float64

The results look logical. Notice how we passed the fahr_to_celsius() function without using the parentheses () after the name of the function. When using apply, you should always leave out the parentheses from the function that you use. Meaning that you should use apply(fahr_to_celsius) instead of apply(fahr_to_celsius()). Why? Because the apply() method will execute and use the function itself in the background when it operates with the data. If we would pass our function with the parentheses, the fahr_to_celsius() function would actually be executed once before the loop with apply() starts (hence becoming unusable), and that is not what we want. Our previous command only returned the Series of temperatures to the screen, but naturally we can also store them permanently into a new column (overwriting the old values):

data["TEMP_C"] = data["TEMP_F"].apply(fahr_to_celsius)

A nice thing with apply() is that we can also apply the function on several columns at once. Below, we also sort the values in descending order based on values in MIN column to see that applying our function really works:

cols = ["TEMP_F", "MIN", "MAX"]
result = data[cols].apply(fahr_to_celsius)
result.sort_values(by="MIN", ascending=False).head()
TEMP_F MIN MAX
196775 25.000000 25.000000 29.444444
154531 25.555556 24.444444 27.777778
188167 25.000000 24.444444 27.777778
188407 23.888889 23.888889 27.777778
188143 24.444444 23.888889 28.888889

You can also directly store the outputs to new columns 'TEMP_C', 'MIN_C', 'MAX_C':

cols = ["TEMP_F", "MIN", "MAX"]
data[cols] = data[cols].apply(fahr_to_celsius)
data.head()
STATION_ID TIME TEMP_F MAX MIN TEMP_C
0 29820 190601010600 1.111111 NaN NaN 1.111111
1 29820 190601011300 0.000000 NaN NaN 0.000000
2 29820 190601012000 -1.111111 NaN NaN -1.111111
3 29820 190601020600 0.555556 NaN NaN 0.555556
4 29820 190601021300 1.666667 NaN NaN 1.666667

In this section, we showed you a few different ways to iterate over rows in pandas and apply functions. The most important thing is that you understand the logic of how loops work and how you can use your own functions to modify the values in a pandas DataFrame. Whenever you need to loop over your data, we recommend using .apply() as it is typically the most efficient one in terms of execution time. However, remember that in most cases you do not actually need to use loops, but you can do calculations in a “vectorized manner” (which is the fastest way) as we learned previously when doing basic calculations in pandas.

String slicing#

We will eventually want to group our data based on month in order to see if the January temperatures in 2020 were higher than on average (which is the goal in our analysis as you might recall). Currently, the date and time information is stored in the column TIME that has a structure yyyyMMddhhmm. This is a typical timestamp format in which yyyy equals to year in four digit format, MM to month (two digits), dd days, hh hours and mm minutes. Let’s have a closer look at the date and time information we have by checking the values in that column, and their data type:

data["TIME"].head()
0    190601010600
1    190601011300
2    190601012000
3    190601020600
4    190601021300
Name: TIME, dtype: int64
data["TIME"].tail()
198329    201910011900
198330    201910012000
198331    201910012100
198332    201910012200
198333    201910012300
Name: TIME, dtype: int64

The TIME column contains several observations per day (and even several observations per hour). The timestamp for the first observation is 190601010600, i.e. from 1st of January 1906 (way back!), and the timestamp for the latest observation is 201910012350. As we can see, the data type (dtype) of our column seems to be int64, i.e. the information is stored as integer values.

We want to aggregate this data on a monthly level. In order to do so, we need to “label” each row of data based on the month when the record was observed. Hence, we need to somehow separate information about the year and month for each row. In practice, we can create a new column (or an index) containing information about the month (including the year, but excluding days, hours and minutes). There are different ways of achieving this, but here we will take advantage of string slicing which means that we convert the date and time information into character strings and “cut” the needed information from the string objects. The other option would be to convert the timestamp values into something called datetime objects, but we will learn about those a bit later. Before further processing, we first want to convert the TIME column as character strings for convenience, stored into a new column TIME_STR:

data["TIME_STR"] = data["TIME"].astype(str)

If we look at the latest time stamp in the data (201910012350), you can see that there is a systematic pattern YEAR-MONTH-DAY-HOUR-MINUTE. Four first characters represent the year, and the following two characters represent month. Because we are interested in understanding monthly averages for different years, we want to slice the year and month values from the timestamp (the first 6 characters), like this:

date = "201910012350"
date[0:6]
'201910'

Based on this information, we can slice the correct range of characters from the TIME_STR column using a specific pandas function designed for Series, called .str.slice(). As parameters, the function has start and stop which you can use to specify the positions where the slicing should start and end:

data["YEAR_MONTH"] = data["TIME_STR"].str.slice(start=0, stop=6)
data.head()
STATION_ID TIME TEMP_F MAX MIN TEMP_C TIME_STR YEAR_MONTH
0 29820 190601010600 1.111111 NaN NaN 1.111111 190601010600 190601
1 29820 190601011300 0.000000 NaN NaN 0.000000 190601011300 190601
2 29820 190601012000 -1.111111 NaN NaN -1.111111 190601012000 190601
3 29820 190601020600 0.555556 NaN NaN 0.555556 190601020600 190601
4 29820 190601021300 1.666667 NaN NaN 1.666667 190601021300 190601

Nice! Now we have “labeled” the rows based on information about day of the year and hour of the day.

Question 3.5#

Create a new column 'MONTH' with information about the month without the year.

Hide code cell content
# Solution

data["MONTH"] = data["TIME_STR"].str.slice(start=4, stop=6)

Grouping and aggregating data#

Basic logic of grouping a DataFrame using .groupby()#

In the following sections, we want to calculate the average temperature for each month in our dataset. Here, we will learn how to use a .groupby() method which is a handy tool for compressing large amounts of data and computing statistics for subgroups. We will use the groupby method to calculate the average temperatures for each month trough these three main steps:

  1. group the data based on year and month using groupby()

  2. calculate the average temperature for each month (i.e. each group)

  3. store the resulting rows into a DataFrame called monthly_data

We have quite a few rows of weather data (N=198334), and several observations per day. Our goal is to create an aggreated DataFrame that would have only one row per month. The .groupby() takes as a parameter the name of the column (or a list of columns) that you want to use as basis for doing the grouping. Let’s start by grouping our data based on unique year and month combination:

grouped = data.groupby("YEAR_MONTH")

Notice, thas it would also be possible to create combinations of years and months “on-the-fly” if you have them in separate columns. In such case, grouping the data could be done as grouped = data.groupby(['YEAR', 'MONTH']). Let’s explore the new variable grouped:

print(type(grouped))
print(len(grouped))
<class 'pandas.core.groupby.generic.DataFrameGroupBy'>
826

We have a new object with type DataFrameGroupBy with 826 groups. In order to understand what just happened, let’s also check the number of unique year and month combinations in our data:

data["YEAR_MONTH"].nunique()
826

Length of the grouped object should be the same as the number of unique values in the column we used for grouping (YEAR_MONTH). For each unique value, there is a group of data. Let’s explore our grouped data further by check the “names” of the groups (five first ones). Here, we access the keys of the groups and convert them to a list so that we can slice and print only a few of those to the sceen:

list(grouped.groups.keys())[:5]
['190601', '190602', '190603', '190604', '190605']

Let’s check the contents for a group representing January 1906. We can get the values for that month from the grouped object using the get_group() method:

# Specify a month (as character string)
month = "190601"

# Select the group
group1 = grouped.get_group(month)
group1
STATION_ID TIME TEMP_F MAX MIN TEMP_C TIME_STR YEAR_MONTH MONTH
0 29820 190601010600 1.111111 NaN NaN 1.111111 190601010600 190601 01
1 29820 190601011300 0.000000 NaN NaN 0.000000 190601011300 190601 01
2 29820 190601012000 -1.111111 NaN NaN -1.111111 190601012000 190601 01
3 29820 190601020600 0.555556 NaN NaN 0.555556 190601020600 190601 01
4 29820 190601021300 1.666667 NaN NaN 1.666667 190601021300 190601 01
... ... ... ... ... ... ... ... ... ...
88 29820 190601301300 -2.222222 NaN NaN -2.222222 190601301300 190601 01
89 29820 190601302000 -6.111111 NaN NaN -6.111111 190601302000 190601 01
90 29820 190601310600 -7.777778 NaN NaN -7.777778 190601310600 190601 01
91 29820 190601311300 -1.111111 NaN NaN -1.111111 190601311300 190601 01
92 29820 190601312000 -0.555556 NaN NaN -0.555556 190601312000 190601 01

93 rows × 9 columns

As we can see, a single group contains a DataFrame with values only for that specific month. Let’s check the DataType of this group:

type(group1)
pandas.core.frame.DataFrame

So, one group is a pandas DataFrame which is really useful, because it allows us to use all the familiar DataFrame methods for calculating statistics etc. for this specific group which we will see shortly. It is also possible to iterate over the groups in our DataFrameGroupBy object which can be useful if you need to conduct and apply some more complicated subtasks for each group. When doing so, it is important to understand that a single group in our DataFrameGroupBy actually contains not only the actual values, but also information about the key that was used to do the grouping. Hence, when iterating we need to assign the key and the values (i.e. the group) into separate variables. Let’s see how we can iterate over the groups and print the key and the data from a single group (again using break to only see what is happening):

# Iterate over groups
for key, group in grouped:
    # Print key and group
    print("Key:\n", key)
    print("\nFirst rows of data in this group:\n", group.head())

    # Stop iteration with break command
    break
Key:
 190601

First rows of data in this group:
    STATION_ID          TIME    TEMP_F  MAX  MIN    TEMP_C      TIME_STR  \
0       29820  190601010600  1.111111  NaN  NaN  1.111111  190601010600   
1       29820  190601011300  0.000000  NaN  NaN  0.000000  190601011300   
2       29820  190601012000 -1.111111  NaN  NaN -1.111111  190601012000   
3       29820  190601020600  0.555556  NaN  NaN  0.555556  190601020600   
4       29820  190601021300  1.666667  NaN  NaN  1.666667  190601021300   

  YEAR_MONTH MONTH  
0     190601    01  
1     190601    01  
2     190601    01  
3     190601    01  
4     190601    01  

Here, we can see that the key contains the name of the group (i.e. the unique value from YEAR_MONTH).

Aggregating data with groupby()#

We can, for example, calculate the average values for all variables using the statistical functions that we have seen already (e.g. mean, std, min, max, median). To calculate the average temperature for each month, we can use the mean() function. Let’s calculate the mean for all the weather related data attributes in our group at once:

# Specify the columns that will be part of the calculation
mean_cols = ["TEMP_F", "TEMP_C"]

# Calculate the mean values all at one go
mean_values = group1[mean_cols].mean()
mean_values
TEMP_F   -0.537634
TEMP_C   -0.537634
dtype: float64

As a result, we get a pandas Series with mean values calculated for all columns in the group. Notice that if you want to convert this Series back into a DataFrame (which can be useful if you e.g. want to merge multiple groups), you can use command .to_frame().T which first converts the Series into a DataFrame and then transposes the order of the axes (the label names becomes the column names):

# Convert to DataFrame
mean_values.to_frame().T
TEMP_F TEMP_C
0 -0.537634 -0.537634

To do a similar aggregation with all the groups in our data, we can actually combine the groupby() function with the aggregation step (such as taking the mean, median etc. of given columns), and finally restructure the resulting DataFrame a bit. This can be at first a bit harder to understand, but this is how you would do the grouping and aggregating the values as follows:

# The columns that we want to aggregate
mean_cols = ["TEMP_F", "TEMP_C"]

# Group and aggregate the data with one line
monthly_data = data.groupby("YEAR_MONTH")[mean_cols].mean().reset_index()
monthly_data
YEAR_MONTH TEMP_F TEMP_C
0 190601 -0.537634 -0.537634
1 190602 -1.044974 -1.044974
2 190603 -2.485066 -2.485066
3 190604 2.740741 2.740741
4 190605 10.722820 10.722820
... ... ... ...
821 201906 14.990715 14.990715
822 201907 17.288769 17.288769
823 201908 17.747080 17.747080
824 201909 13.132371 13.132371
825 201910 8.750000 8.750000

826 rows × 3 columns

As we can see, aggregating the data in this way is fairly straightforward and fast process requiring merely a single command. So what did we actually do here? We i) grouped the data, ii) selected specific columns from the result (mean_cols), iii) calculated the mean for all of the selected columns of the groups, and finally 4) reset the index. Resetting the index at the end is not necessary, but by doing it, we turn the YEAR_MONTH values into a dedicated column in our data (which would be otherwise store as index) .

What might not be obvious from this example is the fact that hidden in the background, each group is actually iterated over and the aggregation step is repeated for each group. For you to better understand what happens, we will next repeat the same process by iterating over groups and eventually creating a DataFrame that will contain the mean values for all those weather attributes that we were interested in. In this approach, we will first iterate over the groups, then calculate the mean values, store the result into a list, and finally merge the aggregated data into a DataFrame called monthly_data.

# Create an empty list for storing the aggregated rows/DataFrames
data_container = []

# The columns that we want to aggregate
mean_cols = ["TEMP_F", "TEMP_C"]

# Iterate over the groups
for key, group in grouped:
    # Calculate mean
    mean_values = group[mean_cols].mean()

    # Add the ´key´ (i.e. the date+time information) into the Series
    mean_values["YEAR_MONTH"] = key

    # Convert the pd.Series into DataFrame and
    # append the aggregated values into a list as a DataFrame
    data_container.append(mean_values.to_frame().T)

# After iterating all groups, merge the list of DataFrames
monthly_data = pd.concat(data_container)
monthly_data
TEMP_F TEMP_C YEAR_MONTH
0 -0.537634 -0.537634 190601
0 -1.044974 -1.044974 190602
0 -2.485066 -2.485066 190603
0 2.740741 2.740741 190604
0 10.72282 10.72282 190605
... ... ... ...
0 14.990715 14.990715 201906
0 17.288769 17.288769 201907
0 17.74708 17.74708 201908
0 13.132371 13.132371 201909
0 8.75 8.75 201910

826 rows × 3 columns

As a result, we get identical results as with the earlier approach that was done with a single line of code (except for the position of the YEAR_MONTH column).

So which approach should you use? From the performance point of view, we recommend using the first approach (i.e. chaining) which does not require you to create a separate for loop, and is highly performant. However, this approach might be a bit more difficult to read and comprehend (the loop might be easier). Also sometimes you want to include additional processing steps inside the loop which can be hard accomplish by chaining everything into a single command. Hence, it is useful to know both of these approaches for doing aggregations with the data.

Case study: Detecting warm months#

Now, we have aggregated our data on monthly level and all we need to do is to check which years had the warmest January temperatures. A simple approach is to select all January values from the data and check which group(s) have the highest mean value. Before doing this, let’s separate the month information from our timestamp following the same approach as previously we did when slicing the year-month combination:

monthly_data["MONTH"] = monthly_data["YEAR_MONTH"].str.slice(start=4, stop=6)
monthly_data.head()
TEMP_F TEMP_C YEAR_MONTH MONTH
0 -0.537634 -0.537634 190601 01
0 -1.044974 -1.044974 190602 02
0 -2.485066 -2.485066 190603 03
0 2.740741 2.740741 190604 04
0 10.72282 10.72282 190605 05

Now we can select the values for January from our data and store it into a new variable january_data. We will also check the highest temperature values by sorting the DataFrame in a descending order:

january_data = monthly_data.loc[monthly_data["MONTH"] == "01"]
january_data.sort_values(by="TEMP_C", ascending=False).head()
TEMP_F TEMP_C YEAR_MONTH MONTH
0 1.302294 1.302294 200801 01
0 1.293952 1.293952 197501 01
0 1.273434 1.273434 198301 01
0 1.15491 1.15491 199201 01
0 1.069869 1.069869 198901 01

By looking at the order of YEAR_MONTH column, we can see that January 2020 indeed was on average the warmest month on record based on weather observations from Finland.

Automating the analysis#

Now we have learned how to aggregate data using pandas. average temperatures for each month based on hourly weather observations. One of the most useful aspects of programming, is the ability to automate processes and repeat analyses such as these for any number of weather stations (assuming the data structure is the same).

Hence, let’s now see how we can repeat the previous data analysis steps for 15 weather stations located in different parts of Finland containing data for five years (2015-2019). The idea is that we will repeat the process for each input file using a (rather long) for loop. We will use the most efficient alternatives of the previously represented approaches, and finally will store the results in a single DataFrame for all stations. We will learn how to manipulate filepaths in Python using the pathlib module and see how we can list our input files in the data directory data/finnish_stations. We will store those paths to a variable file_list, so that we can use the file paths easily in the later steps.

Managing and listing filesystem paths#

In Python there are two commonly used approaches to manage and manipulate filepaths, namely os.path sub-module and a newer pathlib module (available since Python 3.4) which we will demonstrate here. The built-in module pathlib provides many useful functions for interacting and manipulating filepaths on your operating system. On the following, we have data in different sub-folders and we will learn how to use the Path class from the pathlib library to construct filepaths. Next, we will import and use the Path class and see how we can construct a filepath by joining a folder path and file name:

from pathlib import Path

# Initialize the Path
input_folder = Path("data/finnish_stations")

# Join folder path and filename
fp = input_folder / "028360.txt"
fp
PosixPath('data/finnish_stations/028360.txt')

Here, we first initialized the Path object and stored it in variable input_folder by passing a relative path as a string to directory where all our files are located. Then we created a full filepath to file 028360.txt by adding a forward slash (/) character between the folder and the filename which joins them together (easy!). In this case, our end result is something called a PosixPath which is a filesystem path to a given file on Linux or Mac operating systems. If you would run the same commands on Windows machine, the end result would be a WindowsPath. Hence, the output depends on which operating system you are using. However, you do not need to worry about this, because both types of Paths work exactly the same, no matter which operating system you use.

Both the Path object that we stored in input_folder variable and the PosixPath object that we stored in variable fp are actually quite versatile creatures, and we can do many useful things with them. For instance, we can find the parent folder where the file is located, extract the filename from the full path, test whether the file or directory actually exists, find various properties of the file (such as size of the file or creation time), and so on:

fp.parent
PosixPath('data/finnish_stations')
fp.name
'028360.txt'
fp.exists()
True
# File properties
size_in_bytes = fp.stat().st_size
creation_time = fp.stat().st_ctime
modified_time = fp.stat().st_mtime
print(
    f"Size (bytes): {size_in_bytes}\nCreated (seconds since Epoch): {creation_time}\nModified (seconds since Epoch): {modified_time}"
)
Size (bytes): 1347907
Created (seconds since Epoch): 1723456869.6959033
Modified (seconds since Epoch): 1649532851.0

There are also various other methods that you can do with pathlib, such as rename the files (.rename()) or create folders (.mkdir()). You can see all available methods from pathlib documentation [3]. One of the most useful tools in pathlib is the ability to list all files within a given folder by using the method .glob() which also allows you to add specific search criteria for listing only specific files from the directory:

file_list = list(input_folder.glob("0*txt"))

Here, the result is stored into variable file_list as a list. By default, the .glob() produces something called a generator which is a “lazy iterator”, i.e. a special kind of function that allows you to iterate over items like a list, but without actually storing the data in memory. By enclosing the .glob() search functionality with list() we convert this generator into a normal Python list. Note that we’re using the * character as a wildcard, so any filename that starts with 0 and ends with txt will be added to the list of files. We specifically use the number 0 as the starting part for the search criteria to avoid having metadata files included in the list. Let’s take a look what we got as a result:

print("Number of files in the list:", len(file_list))
file_list
Number of files in the list: 15
[PosixPath('data/finnish_stations/029170.txt'),
 PosixPath('data/finnish_stations/028690.txt'),
 PosixPath('data/finnish_stations/029820.txt'),
 PosixPath('data/finnish_stations/029700.txt'),
 PosixPath('data/finnish_stations/028970.txt'),
 PosixPath('data/finnish_stations/029070.txt'),
 PosixPath('data/finnish_stations/029500.txt'),
 PosixPath('data/finnish_stations/029110.txt'),
 PosixPath('data/finnish_stations/028750.txt'),
 PosixPath('data/finnish_stations/029720.txt'),
 PosixPath('data/finnish_stations/029440.txt'),
 PosixPath('data/finnish_stations/028360.txt'),
 PosixPath('data/finnish_stations/029810.txt'),
 PosixPath('data/finnish_stations/029740.txt'),
 PosixPath('data/finnish_stations/029350.txt')]

Iterate over input files and repeat the analysis#

Now, we should have all the relevant file paths in the file_list, and we can loop over the list using a for loop (again we break the loop after first iteration):

for fp in file_list:
    print(fp)
    break
data/finnish_stations/029170.txt

The data that we have sampled is in regular CSV format which we can read easily with pd.read_csv() function:

data = pd.read_csv(fp)
data.head()
USAF YR--MODAHRMN DIR SPD GUS TEMP MAX MIN
0 29170 201501010050 240.0 7.0 NaN 34.0 NaN NaN
1 29170 201501010120 260.0 5.0 NaN 36.0 NaN NaN
2 29170 201501010150 250.0 8.0 NaN 34.0 NaN NaN
3 29170 201501010220 250.0 8.0 NaN 36.0 NaN NaN
4 29170 201501010250 240.0 8.0 NaN 36.0 NaN NaN

Now we have all the file paths to our weather observation datasets in a list, and we can start iterating over them and repeat the analysis steps for each file separately. We keep all the analytical steps inside a loop so that all of them are repeated to different stations. Finally, we will store the warmest January for each station in a list called results using a regular Python’s append() method and merge the list of DataFrames into one by using pd.concat() function:

# A list for storing the result
results = []

# Repeat the analysis steps for each input file:
for fp in file_list:
    # Read the data from CSV file
    data = pd.read_csv(fp)

    # Rename the columns
    new_names = {
        "USAF": "STATION_NUMBER",
        "YR--MODAHRMN": "TIME",
        "TEMP": "TEMP_F",
    }
    data = data.rename(columns=new_names)

    # Print info about the current input file
    # This is useful to understand how the process proceeds
    print(
        f"STATION NUMBER: {data.at[0,'STATION_NUMBER']}\tNUMBER OF OBSERVATIONS: {len(data)}"
    )

    # Create column
    col_name = "TEMP_C"
    data[col_name] = None

    # Convert temperatures from Fahrenheit to Celsius
    data["TEMP_C"] = data["TEMP_F"].apply(fahr_to_celsius)

    # Convert TIME to string
    data["TIME_STR"] = data["TIME"].astype(str)

    # Parse year and month and convert them to numbers
    data["MONTH"] = data["TIME_STR"].str.slice(start=5, stop=6).astype(int)
    data["YEAR"] = data["TIME_STR"].str.slice(start=0, stop=4).astype(int)

    # Extract observations for the months of January
    january = data[data["MONTH"] == 1]

    # Aggregate the data and get mean values
    columns = ["TEMP_F", "TEMP_C", "STATION_NUMBER"]
    monthly_mean = january.groupby(by=["YEAR", "MONTH"])[columns].mean().reset_index()

    # Sort the values and take the warmest January
    warmest = monthly_mean.sort_values(by="TEMP_C", ascending=False).head(1)

    # Add to results
    results.append(warmest)

# Merge all the results into a single DataFrame
results = pd.concat(results)
STATION NUMBER: 29170	NUMBER OF OBSERVATIONS: 120211
STATION NUMBER: 28690	NUMBER OF OBSERVATIONS: 119674
STATION NUMBER: 29820	NUMBER OF OBSERVATIONS: 40264
STATION NUMBER: 29700	NUMBER OF OBSERVATIONS: 120618
STATION NUMBER: 28970	NUMBER OF OBSERVATIONS: 120891
STATION NUMBER: 29070	NUMBER OF OBSERVATIONS: 40473
STATION NUMBER: 29500	NUMBER OF OBSERVATIONS: 40405
STATION NUMBER: 29110	NUMBER OF OBSERVATIONS: 81164
STATION NUMBER: 28750	NUMBER OF OBSERVATIONS: 81127
STATION NUMBER: 29720	NUMBER OF OBSERVATIONS: 81127
STATION NUMBER: 29440	NUMBER OF OBSERVATIONS: 120947
STATION NUMBER: 28360	NUMBER OF OBSERVATIONS: 35442
STATION NUMBER: 29810	NUMBER OF OBSERVATIONS: 35377
STATION NUMBER: 29740	NUMBER OF OBSERVATIONS: 121654
STATION NUMBER: 29350	NUMBER OF OBSERVATIONS: 116220

Awesome! Now we have conducted the same analysis for 15 weather stations in Finland and it did not took too many lines of code! We were able to follow how the process advances with the printed lines of information, i.e. we did some simple logging of the operations. Let’s finally investigate our results:

results
YEAR MONTH TEMP_F TEMP_C STATION_NUMBER
3 2018 1 27.705512 -2.385827 29170.0
3 2018 1 20.855503 -6.191387 28690.0
0 2015 1 38.708724 3.727069 29820.0
0 2015 1 38.052815 3.362675 29700.0
3 2018 1 25.359090 -3.689395 28970.0
2 2017 1 33.014374 0.563541 29070.0
0 2015 1 39.166185 3.981214 29500.0
0 2015 1 32.306713 0.170396 29110.0
0 2015 1 27.514236 -2.492091 28750.0
0 2015 1 34.983254 1.657363 29720.0
0 2015 1 32.062323 0.034624 29440.0
3 2018 1 19.644055 -6.864414 28360.0
0 2015 1 39.864476 4.369154 29810.0
0 2015 1 34.405518 1.336399 29740.0
0 2015 1 28.617084 -1.879398 29350.0

Each row in the results represents the warmest January at given STATION_NUMBER between the years 2015 and 2019. Based on the YEAR column, the warmest January in most of Finland’s weather stations during this five-year period was in 2015. We can confirm this by checking the value counts of the YEAR column:

results["YEAR"].value_counts()
YEAR
2015    10
2018     4
2017     1
Name: count, dtype: int64

Footnotes#