Estimating street-scale NO2 surface concentrations from TROPOMI observations and high-resolution geographic data

Satellite instruments such as TROPOMI are widely used for comprehensive global monitoring of nitrogen dioxide (NO₂). However, existing satellite retrievals only provide column densities (integrated trace gas concentrations) rather than surface concentrations, which limits their direct applicability for human-health studies

Over recent years, numerous machine learning models for the estimation of surface NO₂ have been developed. Such models use NO₂ vertical column densities (VCDs) from TROPOMI and ancillary input variables, such as meteorological data or bottom-up emission inventories, to predict surface NO₂ concentrations learned from in situ measurements. A consistent finding across studied is that land-use data and road networks are particularly helpful predictors, as they are available at street-scale resolutions and strongly linked to local NO₂ levels. However, their high spatial resolution introduces a major technical challenge: Representing square-kilometer-scale areas requires thousands of input data points, rendering many otherwise suitable neural network architectures, such as multilayer perceptrons, impractical to train. Consequently, previous approaches have relied on spatial aggregation methods, for example by computing coarse metrics such as road density at resolutions of 100 m × 100 m or coarser.

We develop a new methodology that processes such high-resolution ancillary data as images at street-scale resolution (~ 10 m × 10 m or finer), including road networks and building footprints from OpenStreetMap, detailed land-use information from the OSM Land-Use catalogue, and NO_x point sources from the European Release and Transfer Register (E-PRTR). A convolutional neural network is used to encode these high-resolution data into latent features. Combined with the TROPOMI NO₂ VCD and other low-resolution inputs, these are then used to estimate surface NO₂ concentrations via a multiplayer perceptron.

This approach is expected to

• improve predictive accuracy compared to models that rely on aggregation
• enable substantially higher horizontal output resolution down to the street scale
• provide a general framework for estimating surface concentrations of pollutants other than NO₂, as well as full diurnal concentration cycles