File:Overall cross-validation performance for estimating all-source PM2.5 and O3 and the accuracy of fire-sourced PM2.5 in validation against Childs et al.webp

English: "a, Table showing the global and continent-specific performance of the machine learning models estimating all-source daily PM2.5 and O3 based on spatial 10-fold and cluster-based cross-validation (CV), in comparison with the performance of the raw GEOS-Chem outputs. Some spatial cluster could be in two continents; this is why the global total cluster number is smaller than the sum of continent-specific cluster number. The unit of the root mean squared error (RMSE) is µg/m3 for both PM2.5 and O3. The continent-specific CV results was calculated by extracting each continent’s CV data from the global spatial 10-fold CV or global cluster-based CV. b, Density scatter plots showing the performance of our estimated fire-sourced PM2.5, GEOS-Chem simulated fire-sourced PM2.5, Childs et al’s estimated smoke PM2.5 in validation against smoke PM2.5 observed by PurpleAir stations. In this analysis, we included 68041 observations during 2016–2019 linked to 2147 PurpleAir stations in the contiguous US based on date, longitude and latitude. The smoke PM2.5 data and PurpleAir data were sourced from Childs et al.17. The within-block R2 could be interpreted as how much spatial variations of smoke PM2.5 within the 2.0° × 2.5° grid box can the model estimates (our estimated fire-sourced PM2.5 or Childs et al’s estimated smoke PM2.5) account for. c, Correlation matrix showing the Pearson correlations between our estimated fire-sourced PM2.5, GEOS-Chem simulated fire-sourced PM2.5, Childs et al’s estimated smoke PM2.5, and the smoke PM2.5 observed by PurpleAir station. The data used for panel c is the same as the data used for panel b."