1. Understand the Air Quality Problem

Duration1h30

Learning Objectives

Analyze:
[BC02] a first, well-justified set of hypotheses about a new dataset, formed through open-ended exploration rather than a fixed checklist

Air Quality in African Cities

Context

Air pollution is one of the leading environmental health risks worldwide. Fine particulate matter (PM2.5) is particularly dangerous because it penetrates deep into the lungs and bloodstream. Ground-level PM2.5 sensors are expensive and unevenly distributed, especially across African cities, so researchers increasingly rely on satellite measurements (from the Sentinel-5P mission) to estimate ground-level pollution where no physical sensor exists. You have been asked to build a first model that predicts PM2.5 concentration from satellite measurements, and to check whether a model trained in one city can be trusted in another city where it has never measured anything.

Data Description

For this activity you will work with a reduced dataset: measurements from two cities, Kampala and Nairobi.

Variable	Meaning
city	City where the measurement was taken (Kampala or Nairobi in this activity)
date	Date of the satellite pass
hour	Hour of the satellite pass (UTC)
site_latitude	Latitude of the monitoring site
site_longitude	Longitude of the monitoring site
pm2_5	Ground-level PM2.5 concentration (µg/m³) — the target variable
sulphurdioxide_so2_column_number_density	Sulphur dioxide (SO2) column density, from satellite
carbonmonoxide_co_column_number_density	Carbon monoxide (CO) column density, from satellite
nitrogendioxide_no2_column_number_density	Nitrogen dioxide (NO2) column density, from satellite
formaldehyde_tropospheric_hcho_column_number_density	Tropospheric formaldehyde (HCHO) column density, from satellite
uvaerosolindex_absorbing_aerosol_index	UV absorbing aerosol index
ozone_o3_column_number_density	Ozone (O3) column density, from satellite
uvaerosollayerheight_aerosol_height	Aerosol layer height
cloud_cloud_fraction	Cloud cover fraction

Understanding CRISP-DM

Data science projects tend to follow a similar overall process, formalized as CRISP-DM (Cross-Industry Standard Process for Data Mining): Business Understanding, Data Understanding, Data Preparation, Modeling, Evaluation, and Deployment. You will not go through all six phases in a single session, but recognizing which one you are in helps you ask the right questions at the right time.

This activity covers the first two:

Business Understanding (“Understanding the problem” below): understanding the problem, the stakeholders, and what a useful model would need to achieve, before touching any data.
Data Understanding (“Notebook” below): exploring the data itself to see what it actually contains, independently of any modeling choice.

Understanding the problem

Files: reflection/session_2/module_1.md

Question 1: What information is already available in the problem statement? What is missing (e.g., how many monitoring stations, over what time period, how PM2.5 was measured on the ground)?

Question 2: Who would use a model that predicts PM2.5 from satellite data, and for what decision? What would make such a model trustworthy to them?

Question 3: pm2_5 is the target variable. Which other variables do you expect to be genuinely informative predictors, and which ones look more like identifiers than physical measurements?

Question 4: Suppose you train a model only on Kampala data. What could go wrong when you apply it to Nairobi? List at least two concrete risks.

Question 5: Describe the expected type, measurement scale, and analytical role of every variable.

Variable	Expected type or scale	Possible analytical role
city
date
hour
site_latitude
site_longitude
pm2_5
sulphurdioxide_so2_column_number_density
carbonmonoxide_co_column_number_density
nitrogendioxide_no2_column_number_density
formaldehyde_tropospheric_hcho_column_number_density
uvaerosolindex_absorbing_aerosol_index
ozone_o3_column_number_density
uvaerosollayerheight_aerosol_height
cloud_cloud_fraction

Notebook

Files: notebooks/session2/01_discovery.ipynb

Get the project running

Download the Air Quality project archive and unzip it inside your course workspace (the same ~/data-science-toolkit workspace you created in Session 1):

Download Air Quality Project

From the project directory, install the base environment and open the notebook in VS Code:

1
2


cd ~/data-science-toolkit/air_quality
uv sync --group dev

Open notebooks/session2/01_discovery.ipynb and select the .venv kernel of this project, exactly as you did for the Sales notebook in Run a Notebook in VS Code.

Explore

There is no fixed answer here: the goal is to build hypotheses, not reusable code. Look in particular at:

How pm2_5 is distributed in each city
Whether there is a visible trend or seasonality over time
Which columns have missing values, and how many, per city
Whether any satellite variable looks correlated with pm2_5 — a quick correlation heatmap is enough for now; Session 3 has a dedicated module to do this analysis properly

Info

Keep a few notes on what you observed — you will reuse them when interpreting the model you build in the next activity.