Random forest meteorological normalisation models for Swiss PM<sub>10</sub> trend analysis

Abstract

Abstract. Meteorological normalisation is a technique which accounts for changes in meteorology over time in an air quality time series. Controlling for such changes helps support robust trend analysis because there is more certainty that the observed trends are due to changes in emissions or chemistry, not changes in meteorology. Predictive random forest models (RF; a decision tree machine learning technique) were grown for 31 air quality monitoring sites in Switzerland using surface meteorological, synoptic scale, boundary layer height, and time variables to explain daily PM 10 concentrations. The RF models were used to calculate meteorologically normalised trends which were formally tested and evaluated using the Theil-Sen estimator. Between 1997 and 2016, significantly decreasing normalised PM 10 trends ranged between −0.09 and −1.16 µg m −3 yr −1 with urban traffic sites experiencing the greatest mean decrease in PM 10 concentrations at −0.77 µg m −3 yr −1 . Similar magnitudes have been reported for normalised PM 10 trends for earlier time periods in Switzerland which indicates PM 10 concentrations are continuing to decrease at similar rates as in the past. The ability for RF models to be interpreted was leveraged using partial dependence plots to explain the observed trends and relevant physical and chemical processes influencing PM 10 concentrations. Notably, two regimes were suggested by the models which cause elevated PM 10 concentrations in Switzerland: one related to poor dispersion conditions and a second resulting from high rates of secondary PM generation in deep, photochemically active boundary layers. The RF meteorological normalisation process was found to be robust, user friendly and simple to implement, and readily interpretable which suggests the technique could be useful in many air quality exploratory data analysis situations.

Mapping and Understanding Patterns of Air Quality Using Satellite Data and Machine Learning

“…The PD of BLH shows that the model is able to reproduce the pattern of decreasing particle concentrations with increasing BLH (Gupta & Christopher, ; Wagner & Schäfer, ).The shape of the full‐year PD of BLH shown in Figure a is similar to that provided by Grange et al () for observations in Switzerland. They found a reduction of 8 g/m for daily PM10 predictions.…”

“…This tendency was reduced by the choice of the least squares loss function as describes in chapter 2.5.3 but likely still continues to affect the model accuracy. The model performance is comparable to similar studies, also in its underestimation of PM (Hu et al, ; Grange et al, ; Stafoggia et al, ; Zhang et al, ). Tenfold random train/test splits were conducted, resulting in 10 models.…”

“…For example, Park et al () used a RF model to predict PM10 in South Korea and found large influence of AOD, day of the year (DOY), wind speed, and solar radiation on the modeled PM concentrations. Similarly, Grange et al () obtained high feature importances for wind speed, back trajectory cluster, DOY, and air temperature, using the RF approach for PM10 measurements in Switzerland. The present study builds upon the approaches applied in these studies but provides a more in‐depth analysis of model‐inherent relationships.…”

“…However, there is indication that very high BLH values ( 2,500 m) influence PM10 concentrations as well. While the former threshold marks the effects of particle accumulation within a shallow boundary layer, the latter threshold could indicate the formation of a deep boundary layer with stimulated formation of secondary aerosols as suggested by Grange et al (). If BLH is between these thresholds, its explanatory power is limited.…”

“…Seasonality was shown to be an important factor in previous studies (Grange et al, 2018). Here, DOY is used as seasonal proxy.…”