Data Processing and Methods

Spatial Processing

Noise rasters were clipped to Zurich, aggregated to districts via zonal statistics in ArcGIS Pro, and to improve the visual appearance of the maps, water bodies were masked visually without affecting the statistical calculations. Small NoData gaps were filled using a 10-pixel setting where appropriate, while larger gaps such as forests and parks remained unaltered. The noise raster was reclassified into three exposure levels: below 55 dB, 55–65 dB, and above 65 dB. For each district, the share of pixels in each class was calculated. These proportions were then used for the classification of the noise exposure layer, with thresholds for the share of areas above 65 dB set to 1–5%, 5–10%, and above 10%.

Socio-demographic datasets were standardised and converted into percentage values to improve readability and comparability across districts. Age structure was calculated as the share of 0–24, 25–64, and 65+ year-olds per district, with the 65+ share classified by quantiles into up to 35%, 35–38%, and above 38%. Welfare benefit ratios were classified using quantile-based breaks, corresponding to 1–2%, 3–4%, and 5–11%. Education was expressed as population shares by attainment level, with an emphasis on the share without post-compulsory education, classified as below 10%, 10–15%, and above 15%. The foreign-resident share was grouped into below 25%, 25–30%, and above 30%. Income was represented using 2023 median taxable income under the basic tariff, classified into below 50, 50–55, and above 55 thousand CHF.

These preprocessing steps enable effective multivariate exploration, drawing on geovisual analytics principles for pattern revelation (Robinson, 2017).

Thresholds of data

The threshold scheme was applied consistently across the mapped variables to support comparability between districts. For the noise layer, pixel counts were converted into percentages of the total noise-covered area within each district, and these values were then categorized into three classes for the above 65 dB pixels. For socio-demographic indicators, fixed class breaks were retained where already established, while variables with a more natural distribution were grouped into quantile-based thirds. This approach balances interpretability with consistency across the bivariate mapping framework.

Bivariate Mapping

Bivariate choropleth maps were created to visualise the relationship between traffic noise exposure and selected socio-demographic indicators at the district level. A 3x3 classification scheme combined the share of area exposed to more than 65 dB with demographic variables, leveraging perceptually effective color blending and visual variables for good covariation detection (Garlandini & Fabrikant, 2009). Dropdown menus allow layer switching to isolate demographics, reducing clutter and supporting easy map exploration, while popups deliver details on-demand, an essential for interactive cartography (Roth, 2013a). Implemented in ArcGIS Online, this highlights high-noise/vulnerability overlaps, showing potential income-noise disparities (Robinson et al., 2017).

Statistical Analysis

Scatterplots were used to explore the relationship between noise exposure and selected socio-demographic indicators at the district level. For each demographic indicator, a linear trendline was added and the coefficient of determination (R²) was used to assess how much variation could be explained by a linear relationship. Approximate Pearson correlation coefficients (r) were derived from the R² values, with the sign determined by the slope of the trendline. Because the analysis is based on a limited number of districts (n=34), the results were interpreted cautiously and used only as exploratory evidence of association, not as proof of causality. The scatterplots were interpreted together with the bivariate maps, which helped place the statistical patterns in their spatial context.

100x100m demographic layer

For the analysis of spatially distinct patterns on a larger scale, a 100x100m bivariate map was created, illustrating noise exposure, and foreign population share. This layer was created by classifying the mean road noise of each grid cell and subsequently overlaying it with the classified foreign population percentage. Upon analysing the spatial distribution of noise pixels, it was clearly visible that most 100x100m pixels fall within noise class 1 (< 55dB), whereas low numbers of pixels fall within noise class 3 (> 65dB). Only such cells that are located nearby major roads exceed the 65dB limit. Although low noise pollution values are a desirable result for the city of Zurich, the distribution is unsuitable for spatial pattern recognition. Hence, for the final 100x100m bivariate map, lower thresholds of 40dB and 50dB were applied to the classes 1, 2, and 3, respectively.

Uncertainties

Several uncertainties should be acknowledged. Not all datasets are equally current, and for most indicators the available resolution is limited to district level, which limits the spatial detail and can hide differences within individual districts, which is a common challenge in geovisual analysis and the interpretation of spatial patterns (MacEachren & Kraak, 2001). Another limitation concerns the income data. The analysis uses taxable income under the basic tax rate, which is not the same as actual household income and should therefore be interpreted as an income proxy rather than a direct measure of living standards. Finally, aggregating raster noise data to areal units can also hide local hotspots and introduce an aggregation bias. Because of the small number of districts, the Pearson correlations and mapped patterns should be treated as exploratory rather than causal.

Map Design

We designed this bivariate map to show the relationship between noise exposure and sociodemographic patterns in Zurich in a single view, so users can immediately see where noisy areas overlap with different social groups (Hildén, 2022). To make the map more usable, we added a drop-down menu that lets viewers switch between demographic layers, which supports focused comparison without cluttering the map with every variable at once (Robinson et al., 2017). In addition, clicking on an area opens a popup with the relevant local data, so the map works not only as an overview but also as an exploratory interface for attribute-level inspection (Robinson et al., 2017).

The color scheme was chosen to balance analytical clarity with visual distinction. Hildén (2022) argues that bivariate choropleth maps work best when the palette is carefully matched to the data and when the number of classes stays limited, because readability drops quickly as complexity increases. We therefore used a restrained class structure and a legend that makes the two dimensions legible at a glance (Hildén, 2022; MacEachren et al., 2012).

We also muted the basemap so the thematic layer stays visually dominant. This follows the cartographic principle that background detail should not compete with the data, and it supports faster selective attention to the mapped relationship (Fabrikant, 2026; Ware, 2013). In visual search terms, the interface reduces distractors and helps users guide attention toward the thematic pattern they are trying to interpret (Wolfe & Horowitz, 2004).

The interactive elements were added to support exploration rather than to replace the map's overall structure. MacEachren et al. (2012) show that visualization works best when symbolization is intuitive but still efficient for comparison tasks, and popups help by revealing precise values only when needed. Similarly, the lecture materials emphasize that map-based decision support depends on display design and human use context, not just on the data itself (Fabrikant, 2026). For that reason, the dropdown and popups were designed to let users move between broad spatial patterns and local details without losing the overall bivariate relationship.

The map is intended as an exploratory tool rather than a causal argument. It helps reveal possible spatial inequalities in exposure to noise, but the map itself does not explain why those patterns exist (Robinson, 2017). That is why we aimed for a design that is clear, accessible, and easy to interpret, while still allowing users to inspect the overlap between variables in detail (Hildén, 2022; MacEachren et al., 2012).