Spatial joins

Goals of this notebook:

  • Based on the countries and cities dataframes, determine for each city the country in which it is located.
  • To solve this problem, we will use the the concept of a 'spatial join' operation: combining information of geospatial datasets based on their spatial relationship.
In [1]:
%matplotlib inline

import pandas as pd
import geopandas

pd.options.display.max_rows = 10
In [2]:
countries = geopandas.read_file("zip://./data/")
cities = geopandas.read_file("zip://./data/")
rivers = geopandas.read_file("zip://./data/")

Recap - joining dataframes

Pandas provides functionality to join or merge dataframes in different ways, see for an overview and for the full documentation.

To illustrate the concept of joining the information of two dataframes with pandas, let's take a small subset of our cities and countries datasets:

In [3]:
cities2 = cities[cities['name'].isin(['Bern', 'Brussels', 'London', 'Paris'])].copy()
cities2['iso_a3'] = ['CHE', 'BEL', 'GBR', 'FRA']
In [4]:
name geometry iso_a3
26 Bern POINT (7.466975462482424 46.91668275866772) CHE
170 Brussels POINT (4.33137074969045 50.83526293533032) BEL
219 London POINT (-0.1186677024759319 51.5019405883275) GBR
235 Paris POINT (2.33138946713035 48.86863878981461) FRA
In [5]:
countries2 = countries[['iso_a3', 'name', 'continent']]
iso_a3 name continent
0 AFG Afghanistan Asia
1 AGO Angola Africa
2 ALB Albania Europe
3 ARE United Arab Emirates Asia
4 ARG Argentina South America

We added a 'iso_a3' column to the cities dataset, indicating a code of the country of the city. This country code is also present in the countries dataset, which allows us to merge those two dataframes based on the common column.

Joining the cities dataframe with countries will transfer extra information about the countries (the full name, the continent) to the cities dataframe, based on a common key:

In [6]:
cities2.merge(countries2, on='iso_a3')
name_x geometry iso_a3 name_y continent
0 Bern POINT (7.466975462482424 46.91668275866772) CHE Switzerland Europe
1 Brussels POINT (4.33137074969045 50.83526293533032) BEL Belgium Europe
2 London POINT (-0.1186677024759319 51.5019405883275) GBR United Kingdom Europe
3 Paris POINT (2.33138946713035 48.86863878981461) FRA France Europe

But, for this illustrative example, we added the common column manually, it is not present in the original dataset. However, we can still know how to join those two datasets based on their spatial coordinates.

Recap - spatial relationships between objects

In the previous notebook 02-spatial-relationships.ipynb, we have seen the notion of spatial relationships between geometry objects: within, contains, intersects, ...

In this case, we know that each of the cities is located within one of the countries, or the other way around that each country can contain multiple cities.

We can test such relationships using the methods we have seen in the previous notebook:

In [7]:
france = countries.loc[countries['name'] == 'France', 'geometry'].squeeze()
In [8]:
0      False
1      False
2      False
3      False
4      False
238    False
239    False
240    False
241    False
242    False
Length: 243, dtype: bool

The above gives us a boolean series, indicating for each point in our cities dataframe whether it is located within the area of France or not.
Because this is a boolean series as result, we can use it to filter the original dataframe to only show those cities that are actually within France:

In [9]:
name geometry
10 Monaco POINT (7.406913173465057 43.73964568785249)
13 Andorra POINT (1.51648596050552 42.5000014435459)
186 Geneva POINT (6.140028034091699 46.21000754707626)
235 Paris POINT (2.33138946713035 48.86863878981461)

We could now repeat the above analysis for each of the countries, and add a column to the cities dataframe indicating this country. However, that would be tedious to do manually, and is also exactly what the spatial join operation provides us.

(note: the above result is incorrect, but this is just because of the coarse-ness of the countries dataset)

Spatial join operation

**SPATIAL JOIN** = *transferring attributes from one layer to another based on their spatial relationship*

Different parts of this operations: * The GeoDataFrame to which we want add information * The GeoDataFrame that contains the information we want to add * The spatial relationship we want to use to match both datasets ('intersects', 'contains', 'within') * The type of join: left or inner join ![](img/illustration-spatial-join.svg)

In this case, we want to join the cities dataframe with the information of the countries dataframe, based on the spatial relationship between both datasets.

We use the geopandas.sjoin function:

In [10]:
joined = geopandas.sjoin(cities, countries, op='within', how='left')
In [11]:
name_left geometry index_right iso_a3 name_right continent pop_est gdp_md_est
0 Vatican City POINT (12.45338654497177 41.90328217996012) 79.0 ITA Italy Europe 6.213780e+07 2221000.0
1 San Marino POINT (12.44177015780014 43.936095834768) 79.0 ITA Italy Europe 6.213780e+07 2221000.0
2 Vaduz POINT (9.516669472907267 47.13372377429357) 9.0 AUT Austria Europe 8.754413e+06 416600.0
3 Lobamba POINT (31.19999710971274 -26.46666746135247) 152.0 SWZ Swaziland Africa 1.467152e+06 11060.0
4 Luxembourg POINT (6.130002806227083 49.61166037912108) 97.0 LUX Luxembourg Europe 5.941300e+05 58740.0
... ... ... ... ... ... ... ... ...
238 Rio de Janeiro POINT (-43.22696665284366 -22.92307731561596) 22.0 BRA Brazil South America 2.073534e+08 3081000.0
239 São Paulo POINT (-46.62696583905523 -23.55673372837896) 22.0 BRA Brazil South America 2.073534e+08 3081000.0
240 Sydney POINT (151.1832339501475 -33.91806510862875) 8.0 AUS Australia Oceania 2.323241e+07 1189000.0
241 Singapore POINT (103.853874819099 1.294979325105942) 111.0 MYS Malaysia Asia 3.138199e+07 863000.0
242 Hong Kong POINT (114.183063458463 22.30692675357551) 30.0 CHN China Asia 1.379303e+09 21140000.0

243 rows × 8 columns

In [12]:
Asia             59
Africa           57
Europe           46
North America    26
South America    14
Oceania           8
Name: continent, dtype: int64

Lets's practice!

We will again use the Paris datasets to do some exercises. Let's start importing them again:

In [13]:
districts = geopandas.read_file("data/paris_districts_utm.geojson")
stations = geopandas.read_file("data/paris_sharing_bike_stations_utm.geojson")
EXERCISE: Make a plot of the density of bike stations by district

  • Determine for each bike station in which district it is located (using a spatial join!). Call the result `joined`.
  • Based on this result, calculate the number of bike stations in each district (e.g. using `groupby` method; you can use the `size` size method to know the size of each group).
    • Make sure the result is a DataFrame called `counts` with the columns 'district_name' and 'n_bike_stations'.
    • To go from a Series to a DataFrame, you can use the `reset_index` or `to_frame` method (both have a `name` keyword to specify a column name for the original Series values.
  • Add those counts to the original `districts` dataframe, creating a new `districts2` dataframe (tip: this is a merge operation).
  • Calculate a new column 'n_bike_stations_by_area'.
  • Make a plot showing the density in bike stations of the districts.

In [14]:
joined = geopandas.sjoin(stations, districts[['district_name', 'geometry']], op='within')
In [15]:
name bike_stands available_bikes geometry index_right district_name
0 14002 - RASPAIL QUINET 44 4 POINT (450804.448740735 5409797.268203795) 52 Montparnasse
143 14112 - FAUBOURG SAINT JACQUES CASSINI 16 0 POINT (451419.446715647 5409421.528587255) 52 Montparnasse
293 14033 - DAGUERRE GASSENDI 38 1 POINT (450708.2275807534 5409406.941172979) 52 Montparnasse
346 14006 - SAINT JACQUES TOMBE ISSOIRE 22 0 POINT (451340.0264470892 5409124.574548723) 52 Montparnasse
429 14111 - DENFERT-ROCHEREAU CASSINI 24 8 POINT (451274.5111513372 5409609.730783217) 52 Montparnasse
In [16]:
# counting the number of stations in each district
counts = joined.groupby('district_name').size()
Amérique           17
Archives            4
Arsenal             7
Arts-et-Metiers     4
Auteuil            21
dtype: int64
In [17]:
# convert the above in a DataFrame with two columns
counts = counts.reset_index(name='n_bike_stations')
district_name n_bike_stations
0 Amérique 17
1 Archives 4
2 Arsenal 7
3 Arts-et-Metiers 4
4 Auteuil 21
In [18]:
# merging those counts back into the districts dataset
districts2 = districts.merge(counts, on='district_name')
id district_name population geometry n_bike_stations
0 1 St-Germain-l'Auxerrois 1672 POLYGON ((451922.1333912524 5411438.484355546,... 4
1 2 Halles 8984 POLYGON ((452278.4194036503 5412160.89282334, ... 13
2 3 Palais-Royal 3195 POLYGON ((451553.8057660239 5412340.522224233,... 6
3 4 Place-Vendôme 3044 POLYGON ((451004.907944323 5412654.094913081, ... 5
4 5 Gaillon 1345 POLYGON ((451328.7522686935 5412991.278156867,... 4
In [19]:
# calculating the relative number of bike stations by area
districts2['n_bike_stations_by_area'] = (
    districts2['n_bike_stations'] / districts2.geometry.area)
In [20]:
ax = districts2.plot(column='n_bike_stations_by_area',
                    figsize=(15, 6))

The overlay operation

In the spatial join operation above, we are not changing the geometries itself. We are not joining geometries, but joining attributes based on a spatial relationship between the geometries. This also means that the geometries need to at least overlap partially.

If you want to create new geometries based on joining (combining) geometries of different dataframes into one new dataframe (eg by taking the intersection of the geometries), you want an overlay operation.

In [21]:
africa = countries[countries['continent'] == 'Africa']
In [22]:
<matplotlib.axes._subplots.AxesSubplot at 0x7f9e890c1f98>
In [23]:
cities['geometry'] = cities.buffer(2)
In [24]:
geopandas.overlay(africa, cities, how='difference').plot()
<matplotlib.axes._subplots.AxesSubplot at 0x7f9e80dbfb00>
* **Spatial join**: transfer attributes from one dataframe to another based on the spatial relationship * **Spatial overlay**: construct new geometries based on spatial operation between both dataframes (and combining attributes of both dataframes)