A panda, gardening. MidJourney 5
The USDA Plant Hardiness Zone Map tells you which plants will survive the winter where you live. It is built from one number: the average annual minimum temperature at a location, in degrees Fahrenheit. Locations whose coldest night averages within the same five degrees get the same zone label, from 1a in interior Alaska to 13b in Puerto Rico.
Gardeners have used it for decades. It is printed on the back of seed packets.
In November 2023 the USDA released a new version, built by the PRISM group at Oregon State from thirty years of weather station records. It was the first update since 2012, and it moved about half the country half a zone warmer. That was reported widely, and mostly reported badly.
Tonight you have both maps, one row per zip code, and every tool you have learned this week. The question is not whether the country got warmer. The question is what you can honestly say about it from these two files, and where each of the obvious answers goes wrong.
USDA Plant Hardiness Zone Maps for 2012 and 2023, prepared by the PRISM Climate Group at Oregon State University, distributed by zip code. Zip code locations from a public zip code database.
Everything you draw tonight is built from these five.
plt.figure(figsize=(w, h)) # the canvas
plt.xlabel(...) ; plt.ylabel(...) ; plt.title(...) # always
plt.tight_layout() # last
sns.scatterplot(data=df, x='col', y='col', hue='col') # one measurement against another
sns.histplot(data=df, x='col') # the shape of one column
sns.barplot(x=series.values, y=series.index) # a grouped Series, as barsAnd these, from earlier in the week:
pd.concat([a, b], ignore_index=True) # the stacking sentence
pd.merge(left, right, left_on='a', right_on='b') # the join sentence
df.pivot_table(index=, columns=, values=) # the pivot sentence
df.groupby('key')['col'].agg(['count', 'mean']) # split, apply, combine
df.sort_values('col', ascending=False).head(n) # the top-N sentence
df['new'] = df['col'].apply(named_function) # the derived-column sentenceCreate a new notebook named EOD_Day7_Hardiness_Zones.ipynb.
Add a title cell:
# Day 7 EOD: Forty Thousand Zip Codes
Date: 09/09/2026import pandas as pd
import matplotlib.pyplot as plt
import seaborn as sns
base = 'https://eds-217-essential-python.github.io/data/'
zones_2012 = pd.read_csv(base + 'hardiness_zones_2012.csv')
zones_2023 = pd.read_csv(base + 'hardiness_zones_2023.csv')
zipcodes = pd.read_csv(base + 'zip_code_database.csv')Two lines of setup tonight use string operations you have not been taught. Both are the same kind of move: taking a column of text and reshaping it into something you can join on or do arithmetic with. Copy them as given.
Padding a zip code. All three files store zip codes as integers, so 01001 arrives as the number 1001 and the leading zero is gone. Two files that lost their zeros in the same way will still match each other, but the moment you print one or compare it to a real zip code you have a problem. .str.zfill(5) zero-fills a string out to five characters:
zones_2023['zipcode'] = zones_2023['zipcode'].astype(str).str.zfill(5)Splitting a range. The trange column holds text like '-10 to -5'. To do arithmetic on the cold end of that range you need the first piece of it, as a number:
zones['trange_min'] = zones['trange'].str.split().str.get(0).astype(int).str.split() cuts the text at every space, giving ['-10', 'to', '-5']. .str.get(0) takes the first piece from each row. .astype(int) turns '-10' into -10.
On Thursday, 4C told you that .str methods chained together were a Day 7 problem. This is Day 7, and this is the chain. Three methods, in a row, each one operating on what the last one handed it, which is exactly the pattern you have been reading all week in .sort_values(...).head(10), because a chain is just a sentence with more than one verb in it.
zones_2023['zipcode'] = zones_2023['zipcode'].astype(str).str.zfill(5)
zones_2012['zipcode'] = zones_2012['zipcode'].astype(str).str.zfill(5)
zipcodes['zip'] = zipcodes['zip'].astype(str).str.zfill(5)Answer each question with code, then write the answer in a markdown cell underneath, in a complete sentence with the numbers in it.
How many rows and columns does each of the three tables have? What are the column names of the two zone tables, and are they the same?
Run .isnull().sum() on both zone tables. Then look at .head() of either one and say, in one sentence, what zone, trange and zonetitle each hold and which of the three is redundant.
The two files do not have the same number of rows. Before you do anything else, write down two different explanations for that which do not involve anybody making a mistake.
Add a year column to each zone table, then stack the two of them into one table called zones, using the stacking sentence. How many rows?
Add trange_min using the line from the Field Note, then confirm it has no nulls and look at its distinct values with .value_counts().sort_index(). In a markdown cell, say what you notice about the spacing of those values and what that means for the phrase βhow much warmerβ.
Compute the mean trange_min for each year and report the difference. This is the headline number, and you are about to spend the rest of the evening finding out how much it is worth.
Merge zones with zipcodes to attach a state, a latitude and a longitude to every row. The key columns are called different things in the two tables. Compare the row count before and after, then run the merge again with how='left' and find out how many rows failed to match and how many distinct zip codes that is.
In a markdown cell: 103 rows out of 80,455 did not match, covering 63 zip codes. Is that a number you would fix, a number you would mention, or a number you would ignore? Say which and why. There is more than one defensible answer.
Filter located to the 2023 rows only, then draw a scatter plot with longitude across the bottom, latitude up the side, and hue='trange_min'. Make the figure twelve inches by seven. Label both axes and title it.
Something is wrong with that figure, and it is not subtle. Describe it in a markdown cell.
Find the culprit. Use the filter sentence on located to pull out every row with a longitude greater than β60, and print its zip code, state, city, latitude and longitude.
In a markdown cell, two or three sentences. That is one zip code, appearing once in each year: two rows out of eighty thousand. Look up roughly where latitude 48.3, longitude β2.1 is. Is the temperature reading wrong, or is something else wrong? And say plainly what those two rows did to your figure.
Nothing about those rows is unusual in a table. They are not null, not duplicated, not miscast, and they would survive every cleaning step you learned on Thursday untouched. .describe() on longitude would have shown a maximum of β2.12 and you would have had no reason to look at it.
A figure found them in one second, because a plot is the only summary of a dataset that gives an outlier as much room as it gives everything else. This is a large part of what visualization is for, and it happens before any of the storytelling.
Build a table called usa holding only the rows with longitude less than β60, and redraw the 2023 map from it. Then draw the same map for 2012.
You have just drawn a recognizable map of the United States without a mapping library, a projection, or a shapefile. In a markdown cell, one sentence on why that worked here. Then one more: run usa['state'].nunique() and say which parts of the country are not in these files at all.
Put the two maps side by side on your screen and try to see the difference between them. In one honest sentence: can you?
Comparing two maps by eye does not work. So compute the change and map that instead, which is the move you learned this morning.
Use the pivot sentence to build a table with one row per zip code and one column per year. Index it on ['zipcode', 'state', 'latitude', 'longitude'] so those come along for the ride, put year across the columns and trange_min in the cells, then .reset_index(). What shape is it, and why is that number smaller than 80,000?
Add a temp_diff column: the 2023 value minus the 2012 value. Count its nulls and say what a null means here. Then drop those rows.
π The two year columns are named with the integers 2012 and 2023, not with strings, because that is what was in the year column you pivoted on. So it is change[2023], with no quotes. Everything else about column-to-column arithmetic is the same as it was on Thursday.
Draw a histogram of temp_diff. Label it. Then print change['temp_diff'].value_counts().sort_index() and read the two together.
In a markdown cell, three or four sentences: what is the most common value, what is the second most common, and how many distinct values are there in total? Given that, is temp_diff a measurement of how much a place warmed, or is it something else? Be precise.
The tails of that histogram are worth a minute. Use the top-N sentence twice to show the five zip codes with the largest increase and the five with the largest decrease, with their states and coordinates.
In a markdown cell, two or three sentences: a zip code that moved by 25 or 30 Β°F has moved five or six whole zones. Do you believe that is a change in the climate? What else could produce it? (Look at where they are.)
temp_diff has too many values to use as a hue= and too few to be interesting as a number. Write a named function called classify_shift that takes one difference and returns 'warmer', 'colder' or 'unchanged', apply it to build a shift column, and count the three categories.
Draw the change map: longitude and latitude again, this time with hue='shift'. Twelve by seven, labelled, titled.
In a markdown cell, four or five sentences. This is the figure the whole evening was for, so describe it properly. Where is the country almost uniformly warmer? Where is it patchy? Where are the 'colder' points, and does their location connect to your answer to question 19?
Then compare this figure with the two maps from question 13, and say in one sentence why the difference had to be computed rather than looked at.
Group change by state and use .agg(['count', 'mean']) on temp_diff. Show the ten states with the largest mean increase.
Read the count column before you read the mean column, the way you have every day since Friday. Two entries in that top ten should stop you. Name them and say why in one sentence each.
Build a Series of the ten largest mean increases, sorted, and draw it as horizontal bars using the .values and .index idiom. Label both axes with units and title it.
Now the other end. Show the five states with the smallest mean increase, with their counts. In a markdown cell, one or two sentences: one of those states has more zip codes in it than all but one other state in the file. Does that make its small number more trustworthy or less, and does it make the ranking more interesting or less?
In a single markdown cell of 250 to 350 words, answer this:
A local newspaper is running a story headlined βOur State Is Now a Zone Warmer.β They have your two files and they want one figure and one number from you. What do you send them, and what do you tell them it does not mean?
Your answer must cite at least three specific numbers you computed tonight, must name at least two distinct reasons the comparison is harder than it sounds, and must end by naming the single figure you would send. Complete sentences, no bullet fragments.
Look back through your notebook. You made six figures tonight, and they did four different jobs:
Only one of the six was the figure the evening was about. The other five were how you found out what you had. That ratio is normal, and it is worth expecting: most of the plots you make in your career will never be shown to anybody, because their job is to tell you something before you say it out loud.
Before you close your notebook, check that:
temp_diff as a measured temperature changeTomorrow you stop being handed datasets. You and your team pick your own, and walk it through all ten steps yourselves. Bring the notebook you started this afternoon.