Population-weighted aggregation

When aggregating gridded climate data to administrative regions, a simple spatial average can be misleading if population is unevenly distributed across the region. Population weighting provides a more accurate measure of the conditions experienced by people, which is particularly important for health-related analyses.

Why use population weighting?¶

Consider a region where most people live in urban areas concentrated in one part of the territory, while the rest is sparsely populated. An unweighted spatial average would give equal importance to all grid cells, regardless of whether people actually live there. This means the resulting average may not accurately represent the conditions experienced by the typical person in that region.

Population weighting solves this problem by giving more weight to grid cells where more people live. This ensures that the aggregated climate values better reflect the actual exposure of the population.

Population-weighted aggregation is especially valuable when:

• Analyzing health impacts of climate and environmental conditions

• Population is concentrated in specific areas (cities, valleys, coastal zones)

• The measured condition varies significantly across the region

• Results will inform public health decisions or interventions

The population weighting method¶

The basic formula for population-weighted aggregation is:

In practice, this means:

1. Align climate and population data to the same grid

2. Multiply each grid cell varaible by the corresponding population count

3. Sum these weighted values across the region

4. Divide by the total population of the region

In this tutorial, we’ll demonstrate this workflow using PM2.5 air quality data and population data for two districts in Sri Lanka.

Import required libraries¶

We’ll use several Python libraries for this analysis:

• xarray: For working with multi-dimensional climate data

• geopandas: For handling geographic boundaries

• earthkit.transform: For spatial aggregation functions

• numpy: For numerical operations

Load PM2.5 air quality data¶

We start by loading daily PM2.5 (particulate matter) air quality data for Sri Lanka from December 2022 through January 2023. To keep the dataset lightweight, it has been spatially restricted to the Colombo and Kalutara districts rather than the full country. Please note that the data is only included to demonstrate the population weighting method, and should be not be used to analyse PM2.5 levels.

This NetCDF file contains gridded PM2.5 concentration values across space and time:

Load district boundaries¶

Next, we load the geographic boundaries for the two districts covered by the PM2.5 dataset, Colombo and Kalutara:

The districts were downloaded from Humanitarian Data Exchange (HDX).

Select data for one year¶

For this example, we’ll focus on PM2.5 data from 2023. We can easily filter the dataset by year using xarray’s datetime indexing:

Load population data¶

Now we load the population raster for Sri Lanka. This file contains population counts per grid cell at approximately 1 km resolution. Note that we rename the coordinate dimensions to match our PM2.5 data (lon and lat):

Align population data to climate grid¶

Critical step: The population and PM2.5 data have different grid alignments. We need to resample the population data to match the PM2.5 grid exactly. We use nearest-neighbor interpolation to preserve population counts:

Calculate weighted PM2.5 values¶

Now we multiply each PM2.5 value by the corresponding population count. This gives us the “population-weighted PM2.5” for each grid cell:

Aggregate to districts¶

Finally, we complete the population-weighted aggregation:

1. Sum weighted PM2.5 (numerator): Total of all PM2.5 × population values within each district

2. Sum population (denominator): Total population within each district

3. Divide: This gives us the population-weighted average PM2.5 for each district

We use earthkit’s spatial.reduce function to perform spatial aggregation for each district polygon:

The resulting dataframe shows population-weighted PM2.5 values for each district and time period. These values represent the average PM2.5 exposure experienced by people in each district, accounting for where people actually live.

Processing multiple years¶

If you have data spanning multiple years, you’ll want to use the appropriate population dataset for each year, since population distributions change over time. Here’s an example that loops through multiple years: