GoKawiilDataset and database construction
We compiled an extensive database of daily average measurements of PM 10 elements related to dust (Al, Ti, Si, Ca and Fe) from all over Europe, to estimate dust concentrations. The bulk of the data was compiled ad hoc for this study, whereas some were extracted from the EBAS database54. The most frequently measured metal that is prevalent in mineral dust was aluminium; measurement data for other dust metals found purely in dust such as silicon and titanium were much scarcer. By contrast, local sources such as road dust as well as non-exhaust traffic and construction emissions affect calcium and iron, making them unsuitable for estimating transported dust in this study40. Metal concentrations were measured with both offline filter-based techniques (inductively coupled plasma (ICP), particle-induced X-ray emission (PIXE) and offline X-ray fluorescence (XRF))55 and online techniques (online XRF X-ray induced acoustic computed tomography (XACT))55,56. Al measurements from these techniques are directly comparable for the purposes of trend analysis, as numerous inter-laboratory comparison studies have shown that these techniques yield highly consistent results for particulate elemental concentrations57,58,59,60,61. Also, grouping model RF cross-validation results by training technique (XRF-based, ICP-based or PIXE; Supplementary Fig. 13) does not indicate any inherent biases between the techniques. Great care was taken in curating the database by flagging data below detection limits and working with data providers to ensure high data quality. After data were aggregated from all providers, various methods were used to ensure data quality. Low concentration values at the detection limit were discarded. To that end, concentration values below a detection cut-off of Al = 0.003 µg m−3 were removed. Also, values that were repeated more than three times in sequence or repeated more than five times (at a third decimal precision) were interpreted as data that were replaced by the detection limit. Further data with a time resolution >1 day were discarded, whereas hourly data were aggregated to daily averages.
Uncertainty of elemental ratios and dust estimate
We quantified the impact of uncertainty in the elemental ratios—defined here as one standard deviation from the bootstrap results shown in Fig. 1b (Si:Al 2.610 ± 0.033, Ti:Al 0.068 ± 0.003, Ca:Al 1.580 ± 0.099 and Fe:Al 0.890 ± 0.052)—on the dust estimates using uncertainty propagation based on the formula: 2.2Al + 2.49Si + 1.63Ca + 1.94Ti + 2.42Fe. This results in a total dust estimate of (13.5 ± 0.377) × Al. Notably, the largest contributions to the overall uncertainty originate from Fe and Ca, which are also influenced by substantial local sources. Zero-intercept fits were used for element–element plots (for example, Ca concentration versus Al concentration) to estimate dust phase mass-based elemental ratios under the assumption that—after blank correction and removal of non-dust contributions—both analytes originate from the same dust endmember. In that case, when the dust contribution vanishes (Al → 0), the co-emitted element should also vanish (Ca → 0), implying a physically meaningful intercept of zero. Also, we tested a linear model with an intercept using bootstrap analysis and found that zero lies within one standard deviation of the mean intercept (intercept ± σ intercept ), supporting the use of a zero-intercept model. Given this result, along with the physical consistency and practical advantages of a zero-intercept regression—such as enabling the use of average ratios to estimate other components—we opted for the zero-intercept model. A summary of the relevant statistics is provided in Supplementary Table 2.
Elemental ratios on days with low amounts of transported dust can be substantially affected by local sources. With a rather high aluminium concentration threshold of >1 μg m−3 for determining the transported dust elemental ratios, we aimed at minimizing the impact of local sources on the elemental ratios. Also, we performed a sensitivity analysis varying the threshold, showing that the elemental ratios vary with cut-offs above 0.5 μg m−3 for Ca:Al and 0.75 μg m−3 for Fe:Al within one standard deviation of the value obtained with a cut-off of 1 μg m−3. Further, we investigated model-driven criteria for dust episodes instead of using a static aluminium concentration cut-off using the DREAM and optical depth values and only including data points that were higher than the 90th percentile in those (each alone or both at the same time), as a proxy for dust events. This approach resulted in similar values (Fig. 1b and Supplementary Fig. 14).
The uncertainty σ mean in modelled yearly mean dust concentrations dust mean (including exceedance and background concentrations) is estimated by accounting for two main components: (1) the uncertainty in the derived dust:Al ratio (RE ratio = 0.377/13.5 = 2.7%) and (2) the relative root mean square error of the modelled annual mean dust concentrations (RRMSE annual mean = 0.45; Supplementary Fig. 3a), as determined through a leave-one-out performance analysis against chemically derived dust concentrations from in situ measurements (equation(1)).
$${\sigma }_{{\rm{mean}}}=\sqrt{{({{\rm{RRMSE}}}_{\text{annual mean}}\times {{\rm{dust}}}_{{\rm{mean}}})}^{2}+{({{\rm{RE}}}_{{\rm{ratio}}}\times {{\rm{dust}}}_{{\rm{mean}}})}^{2}}=\sqrt{{({{\rm{RRMSE}}}_{\text{annual mean}})}^{2}+{{\rm{RE}}}_{{\rm{ratio}}}^{2}}\times {{\rm{dust}}}_{{\rm{mean}}}$$ (1)
The robustness of dust concentration trends over the 10-year study period was evaluated by bootstrapping the linear regression of annual dust concentration time series at each grid cell (100 bootstrap resamples). The reported trend t dust represents the mean of the bootstrap-derived slopes, in which grid cells where the interquartile range includes zero are masked in white (zero trends are assumed for regional means in such cases). The overall uncertainty σ trend,conc in the trend (equation (2)) includes both uncertainty in the dust:Al ratio and the bootstrap-based variability in fitting the trend (RE boot , domain mean relative uncertainty excluding grid cells with non-significant trends), whereas the precision in the yearly dust prediction is already accounted for by the latter (see equation (1)):
$${\sigma }_{{\rm{trend,conc}}}=\sqrt{{({{\rm{R}}{\rm{E}}}_{{\rm{r}}{\rm{a}}{\rm{t}}{\rm{i}}{\rm{o}}}\times {t}_{{\rm{d}}{\rm{u}}{\rm{s}}{\rm{t}}})}^{2}+{({{\rm{R}}{\rm{E}}}_{{\rm{b}}{\rm{o}}{\rm{o}}{\rm{t}}}\times {t}_{{\rm{d}}{\rm{u}}{\rm{s}}{\rm{t}}})}^{2}}=\sqrt{{({{\rm{R}}{\rm{E}}}_{{\rm{r}}{\rm{a}}{\rm{t}}{\rm{i}}{\rm{o}}})}^{2}+{({{\rm{R}}{\rm{E}}}_{{\rm{b}}{\rm{o}}{\rm{o}}{\rm{t}}})}^{2}}\times {t}_{{\rm{d}}{\rm{u}}{\rm{s}}{\rm{t}}},$$ (2)
with \(\sqrt{{({{\rm{RE}}}_{{\rm{ratio}}})}^{2}+{({{\rm{RE}}}_{{\rm{boot}}})}^{2}}\) equal to 0.5, 0.23 and 0.4 for the trend in mean, exceedance and non-exceedance concentrations, respectively.
... continue reading