Data Science: Interpolate Air Quality Measurements with Python
How do we measure air quality? There are a few techniques, but the most popular is to use stationary sensors. We place air quality sensors at specific coordinates; thus, our readings have four parameters [
value]. We can show
values at concrete locations but only as a scatterplot. Moreover, sensors do not cover a big area. We can’t expect observations to be taken at a regular grid, which is easy to interpolate. The final image from the raw data source may look like this one:
This type of point map may be helpful for experts, but the general public will be bored, or much worse – confused – by this kind of visualization. There are so many white areas! Fortunately for us, we can interpolate air pollution in every hex within this map. We are going to learn how to do it step-by-step. We will use the Ordinary Kriging method implemented in
We are going to use two datasets. One represents point measurements of PM2.5 concentrations, and another is a hex-grid over which we will interpolate PM2.5 concentrations. You should be able to follow each step in this article with your data. Just have in mind those constraints:
- both datasets should have the same projection,
- points should be within the hexgrid (or other blocks / MultiPolygon) area.
The safest method of setup is to create a conda environment and install the package within it.
Pyinterpolate works under Python 3.7 and within UNIX systems, so if you are a Windows user, you should consider using Windows Subsystem for Linux (WSL).
Let’s start from
conda environment preparation:
conda create -n [NAME OF YOUR ENV]
Name your environment however you want to. For the sake of this tutorial, we’re going to name it
conda create -n spatial-interpolate
Next, activate your environment:
conda activate spatial-interpolate
Then install core packages:
(spatial-interpolate) conda install -c conda-forge python=3.7 pip notebook
(spatial-interpolate) pip install pyinterpolate
Sometimes you may get an error related to a missing package named libspatialindex_c.so. To install it on Linux, write in terminal:
sudo apt install libspatialindex-dev
And to install it on macOS, use this command in the terminal:
brew install spatialindex
At this point, we should be able to move forward. But, if you encounter any problems, don’t hesitate to ask for help here.
Our task is to transform point observations into a country-wide hex-grid. Observations are stored in a
csv file. Each column is a timestamp, and rows represent stations. Two additional columns
latitude of a station. The hex-grid is a
shapefile which consists of six different types of files. We need to read only one of those – with
.shp prefix – but all others must be placed in the same directory, with the same suffix name. The
csv file could be read with
Shapefile must be loaded with
We will import all functions and packages in the first cell of a notebook, then we read files and check how they look inside.
import pandas as pd import geopandas as gpd import numpy as np from pyproj import CRS from pyinterpolate.semivariance import calculate_semivariance from pyinterpolate.semivariance import TheoreticalSemivariogram from pyinterpolate.kriging import Krige
We have imported a lot of stuff here! Let’s check for what are those packages and functions responsible:
shapefileand transform geometries,
numpy: algebra and matrix operations,
pyproj.CRS: sets Coordinate Reference System (projection of points),
calculate_semivariance: calculate how unequal are observations in the function of distance between them,
TheoreticalSemivariogram: create model of unequality of samples in the function of distance between them,
Krige: Interpolate missing values from given set of points and modeled semivariogram.
Don’t worry if you don’t understand all terms above. We will see what they do in practice in the following steps.
readings = pd.read_csv('pm25_year_2021_daily.csv', index_col='name') canvas = gpd.read_file('hexgrid.shp') canvas['points'] = canvas.centroid
readings is a
DataFrame with 153 columns from which two columns are spatial coordinates. The rest are a daily average of PM 2.5 concentration. It is a time series, but we are going to select one date for further analysis. Variable
canvas represents set of geometries and their id’s. Geometries are stored as hexagonal blocks. We need to create a point grid from those blocks. It is straightforward in
GeoPandas. We can use the
.centroid attribute to get the centroid of each geometry.
We will limit the number of columns in
['01.01.2021', 'x', 'y'] and rename the first column to
df = readings[['01.01.2021', 'x', 'y']] df.columns = ['pm2.5', 'x', 'y']
We can transform our
GeoDataFrame to have the ability to show data as a map. There are two basic steps to do it correctly. First, we should prepare Coordinate Reference System (CRS). It is not simple to assign concrete CRS from scratch because it has multiple properties. Let’s take a look into CRS used to work with geographical data over Poland:
<Projected CRS: EPSG:2180> Name: ETRF2000-PL / CS92 Axis Info [cartesian]: - x[north]: Northing (metre) - y[east]: Easting (metre)for the European Petroleum Survey Group Area of Use: - name: Poland - onshore and offshore. - bounds: (14.14, 49.0, 24.15, 55.93) Coordinate Operation: - name: Poland CS92 - method: Transverse Mercator Datum: ETRF2000 Poland - Ellipsoid: GRS 1980 - Prime Meridian: Greenwich
We are not going to write such long descriptions. Fortunately for us, the European Petroleum Survey Group (EPSG) created a database with numbered indexes that points directly to the most common CRS. We use EPSG:2180 (look into the first row of CRS, and you’ll see it). Python’s
pyproj package has a built-in database of EPSG-CRS pairs. We can provide EPSG number to get specific CRS:
DATA_CRS = CRS.from_epsg('2180')
We have our projection. Bear in mind that we already know CRS/EPSG of a data source. Sometimes it is not so simple with data retrieved from CSV files. You will be forced to dig deeper or try a few projections blindly to get a valid geometry (my first try is usually EPSG 4326 – from GPS coordinates). The next step is to transform our floating-point coordinates into
Point() geometry. After this transformation and setting a CRS, we get a
df['geometry'] = gpd.points_from_xy(df['x'], df['y']) gdf = gpd.GeoDataFrame(df, geometry='geometry', crs=DATA_CRS)
We can show our data on a map! First, we remove
NaN rows, and then we will visualize available points over a canvas used for interpolation.
# Remove NaNs gdf = gdf.dropna()
base = canvas.plot(color='white', edgecolor='black', alpha=0.1, figsize=(12, 12)) gdf.plot(ax=base, column='pm2.5', legend=True, vmin=0, vmax=50, cmap='RdYlGn_r') base.set_title('2021-01-01 Daily PM 2.5 concentrations in Poland');
Build a Variogram Model
Before interpolation, we must create the (semi)variogram model. What is it? It is a function that measures dissimilarity between points in the space. Its foundation is an assumption that close points tend to be similar and distant points are usually different. This simple technique is a potent tool of analysis. To make it work, we must set a few constants:
STEP_SIZE: the distance within points are grouped together. We start from a distance zero, then
0 + STEP_SIZEfrom a point and so on. We test the average dissimilarity of a point and its neighbors within the step size i and step size i+1.
MAX_RANGE: maximum range of analysis. At some range there is no correlation between points and it is a waste of resources to analyze such distant radius from the initial point.
MAX_RANGEshould be smaller than the maximum distance between points in dataset. Usually it is a half of the maximum distance.
The critical thing to notice is to set STEP_SIZE and MAX_RANGE in the units of data. Sometimes you get coordinates in degrees and sometimes in meters. You should always be aware of this fact. Our projection unit is a meter. We can check that the max distance between points could be close to 500 km, thus we set our maximum range to 300 kilometers.
km_c = 10**3 # 1 km is 1000 m STEP_SIZE = km_c * 30 # 30 km MAX_RANGE = km_c * 300 # 300 km
Pyinterpolate operates on
numpy arrays. That’s why we must provide only [
pm2.5] as an array of values instead of
calculate_semivariance() function. We use this array along with
MAX_RANGE and calculate spatial dissimilarity between points with
arr = gdf[['x', 'y', 'pm2.5']].values semivar = calculate_semivariance(arr, step_size=STEP_SIZE, max_range=MAX_RANGE)
It is an intermediate step. If you print
semivar you will see an array with three columns:
array([[0.00000000e+00, 5.05416667e+00, 6.00000000e+01], [3.00000000e+04, 5.76951786e+01, 5.60000000e+01], [6.00000000e+04, 9.47646939e+01, 9.80000000e+01], [9.00000000e+04, 7.46993333e+01, 1.50000000e+02], [1.20000000e+05, 7.45332632e+01, 1.90000000e+02], [1.50000000e+05, 1.31586140e+02, 2.28000000e+02], [1.80000000e+05, 1.24392830e+02, 2.12000000e+02], [2.10000000e+05, 1.40270976e+02, 2.46000000e+02], [2.40000000e+05, 1.41650229e+02, 2.18000000e+02], [2.70000000e+05, 1.42644732e+02, 2.24000000e+02]])
What is the meaning of each column?
- Column 1 is the lag. We test all points within
(lag - 0.5 * step size, lag + 0.5 * step size]. (It is done under the hood with function
select_values_in_range()).. To make it easier to understand let’s look into figure below that shows lags, step size and max range from the perspective of a single point.
- Column 2 is the semivariance. We can understand this parameter as a measurement of dissimilarity between points within a specific ranges given by a lag and a step size. It is averaged value for all points and their neighbors per lag.
- Column 3 is a number of points within specific distance.
What we have created is the experimental semivariogram. With it, we can create a theoretical model of semivariance. Pyinterpolate has a particular class for semivariogram modeling named
TheoreticalSemivariogram. During initialization, it takes an array of known points and the experimental semivariogram. We can model semivariogram manually, but
TheoreticalSemivariogram class has a method to do it automatically. It is named
.find_optimal_model(). If we set
verbose parameter to
True during initialization, we see how different theoretical models behave during tests. The last helpful method is
.show_semivariogram() to observe the difference between empirical semivariogram and created theoretical curve.
ts = TheoreticalSemivariogram(points_array=arr, empirical_semivariance=semivar, verbose=True) ts.find_optimal_model() ts.show_semivariogram()
As you see, the dissimilarity increases over distance. Distant points tend to have different PM 2.5 concentrations.
Interpolate Missing Values with Kriging
The last step of our tutorial is an interpolation. We use the Kriging method. It is a particular type of function which predicts values at unknown locations based on the closest neighbors and theoretical semivariance. In practice, it multiplies concentrations of n-closest neighbors by weighs related directly to the distance from the predicted location to its neighbor.
Pyinterpolate makes this operation extremely simple. The small problem is understanding the API of the package. We can perform prediction for a single coordinate but the function returns a list of values where only the first element is estimated output and other elements are related to the deeper analysis (this is a significant part, we will uncover it in another tutorial).
Kriging is initialized by
Krige class which takes as input
TheoreticalSemivariogram and array with points with known values:
pk = Krige(ts, arr)
Do you remember when we have created centroids for our hexagonal canvas? We have named column with these
points. Now we use every row of this column and interpolate PM 2.5 concentration for it! We will be working with
GeoDataFrame , and it allows us to use
apply method that is used with
def get_prediction(upts, mdl, nn=8): pt = [upts.x, upts.y] predicted = mdl.ordinary_kriging(pt, nn) return pd.Series([upts, predicted, predicted]) # Predict predicted = canvas['points'].apply(get_prediction, mdl=pk) predicted.columns = ['coordinates', 'yhat', 'error'] # Merge with canvas cdf = canvas.join(predicted) cdf = cdf[['geometry', 'yhat', 'error']]
get_prediction() is passed in the
apply method. It takes three parameters:
upts: unknown point as a
mdl: Kriging model,
nn: number of closest neighbors to interpolate missing value, we set it to default 8.
We predict unknown point values with
.ordinary_kriging() method. There are other types of Kriging within the
Krige class; that’s why package’s maintainers didn’t name the prediction method as
predict(). Ordinary Kriging returns four objects:
- Estimated value.
- Estimated (local) mean.
We have only interest in the first element from a list and optionally the second. Weights and estimated mean are essential for debugging and a close investigation of the results.
Below the body of a
get_prediction() function, we have a classical operation on
DataFrames. Finally, we can show results:
base = canvas.plot(color='white', edgecolor='black', alpha=0.1, figsize=(12, 12)) cdf.plot(ax=base, column='yhat', legend=True, vmin=0, vmax=50, cmap='RdYlGn_r') base.set_title('Interpolated 2021-01-01 Daily PM 2.5 concentrations in Poland');
Isn’t it much more interesting than scatterplot?
- More about Semivariogram Modeling within
- More about Kriging within
- Learn about spatial interpolation!
- Try to reproduce this tutorial with other measurements. For example: temperature readings, height above sea level, humidity…