Exponentially Smoothed Hourly PM2.5 Timeseries Maps, 100m, Metropolitan Area Stuttgart, Germany, 2023

Description Disclaimer: This is a prototype version, which contains data for January 1-10, 2023. The temporal extent will soon be updated to the entire year. Introduction This repository contains spatial datasets describing atmospheric particulate matter 2.5 (PM 2.5) concentration at the surface level. The spatial extent covers the metropolitan area of Stuttgart, Germany and includes parts of the northern Black Forest to the west (bounding box: 48-49°N, 8-10°E). In the temporal dimension the data is provided in hourly steps for the entire year of 2023. Maps were generated with a spatial prediction method that combines open low-cost sensor data from Citizen Science / Civic Tech platforms Sensor.Community, OpenSenseMap, and PurpleAir. All maps are results generated within the Open Earth Monitor & Cyberinfrastructure (OEMC) project and its use case on Air Quality Monitoring at Regional Scale. Stakeholders in this use case Sensor.Community and OpenSenseMap, which are two Civic Tech / Citizen Science platforms that collect, display and host real-time air quality measurements from Low-Cost-Sensor networks. The objective of the use case was to (i) implement a spatial prediction method that leverages the dense timeseries data using (ii) open source software to (iii) produce PM2.5 map at high temporal and spatial resolutions. Details Method Data in the study area shows a high density in both the temporal dimension (hourly) and the spatial dimension (~700 sensors; ~0.175 /km²). To leverage these properties we implemented a spatial prediction method that boroughs concepts from the field of statistical forecasting. Exponential smoothing uses "weighted averages of past observations, with the weights decaying exponentially as the observations get older" (Hyndman et al. 2021). The method benfits from the temporal autocorrelation in the measurement data. In an exploratory analysis we found a suitable value for smoothing parameter alpha of 0.8, which give a heavier weight to more recent observations. To model PM2.5, we create multiple temporal PM2.5 derivatives including the mentioned exponentially smoothed variable, the 1-hour lag, the 24-hour lag, and the 8-hour rolling average. They are transferred to the spatial domain of other covariates through nearest-neighbor interpolation. The prediction is supported by additional temporally dynamic covariates from the climate model ERA5-Land reanalysis (temperature, relative humidity, wind direction, wind speed) and from the Copernicus atmospheric monitoring system (CAMS; surface-level PM2.5). Several static covariates support the model in differentiating between urban and rural areas (elevation, population, traffic network, natural areas). We use a Universal Kriging method for map prediction, which has proven useful for air quality applications. A relatively simple Random Forests model predicts the trend component. Model residuals are then interpolated in space through Kriging. The sum of the two map layers forms the final prediction map. Validation is performed with indenpent official EEA measurement station data. These represent the gold standard in terms of air quality measurement accuracy. Input Data & Code Variable Short Name Type Original Resolution Resampling to 100 m PM2.5 Exponentially Smooted pm25_smoothed dynamic points nearest neighbor interpolation PM2.5 Time Lag 1 Hour pm25_lag1 dynamic points nearest neighbor interpolation PM2.5 Time Lag 24 Hours pm25_lag24 dynamic points nearest neighbor interpolation PM2.5 8-hour Rolling Average pm25_roll_8h dynamic points nearest neighbor interpolation ERA5 Relative Humidity rh dynamic ~ 9km bilinear resampling ERA5 Temperature t2m dynamic ~ 9km bilinear resampling ERA5 Wind Speed ws dynamic ~ 9km bilinear resampling ERA5 Wind Direction wd dynamic ~ 9km bilinear resampling CAMS PM2.5 cams_pm2p5 dynamic ~ 9km bilinear resampling Elevation Elevation static 25m bilinear resampling Population Population static 1km bilinear resampling Highway Highway static 1km bilinear resampling Major_Roads Major_Roads static 1km bilinear resampling Minor_Roads Minor_Roads static 1km bilinear resampling Corine Land Cover Natural Areas CLC_Nat static 100m none Data Format Maps are shipped as Zarr stores, an efficient data format particularily suited for multidimensional data. For better compatibility with Zenodo these archives are zipped. Check the supplementary Python notebook for instructions on how to open Zarr stores.