Using machine learning regression to model ambient ultrafine particle concentrations along a runway trajectory near a major airport

Background: Ultrafine particles (UFP; <100 nm diameter) are of concern due to their adverse health impacts. UFP are highly variable in space and time, given their intermittent source emissions, formation, and removal processes, leading to challenges in accurate characterization of source contributions. Recent studies have shown that airports are contributors to local air pollution, but research is needed to understand the impact from commercial aircraft during landing and take-off and how far from flight paths these impacts are observed. Our aim was to use machine-learning regression modeling approaches to better understand ground-based UFP contributions from in-flight aircraft during landings at Boston Logan International Airport (MA, USA) along a main runway trajectory.

Methods: We measured particle number concentration (PNC; a proxy for UFP) and meteorology from April to September 2017 at six sites that were varying distances from an arrival flight path. Concurrent flight activity data were obtained from the U.S. Federal Aviation Administration. We aggregated PNC by hour using the mean, median and 95^th percentile of 1-sec PNC. The natural log-transformed PNC [ln(PNC)], spatial-temporal covariates were included such as flight activity, meteorology, and weekday/weekend. We used random forest (RF) regression to identify key covariates and optimize prediction of PNC. Each tree was grown by a bootstrap sample with random subsets of predictors selected at each split. Final models were based on the average results of all trees and outputs were compared to generalized linear regression models (GLM).

Results: RF models of 1-hr median ln(PNC) had site specific R² between 54% and 67% across the six monitoring sites. While meteorological variables were ranked most important in all models, flight frequency (of aircraft landing on the 4L/R runways) increased in model importance when comparing the 95^th percentile models to the 50^th percentile at near-airport sites. Similar trends were observed at backgrounds sites, even at a site >15 km away from 4L/R runways. Arrival flight frequency explained more variance at near-airport sites compared to background sites and increased in model importance when comparing the 95^th percentile to the median. The RF models explained more variance using the covariates in site specific GLM models.

Conclusions: Our results suggest that arrival aircraft contribute intermittently but significantly to UFP concentrations near flight paths in closer proximity to the airport. Collection of real-time PNC and flight activity data enhance the ability to quantify contributions from aircraft, and understand how these aircraft contribute to community-level PNC and understand the spatial extent of these exposures.