Portal de Eventos, Congreso Colombiano y Conferencia Internacional de Calidad del Aire y Salud Pública

Tamaño de la fuente: 
Analysis of Pollutant Dynamics in a High Andean City using satellite data, a Lagrangian approach and Statistical Learning Techniques: Classification Tree and Random Forest
Juan Felipe Mendez Espinosa, Ricardo Morales Betancourt, Luis Carlos Belalcazar Ceron

Última modificación: 11/06/2019

Resumen


Understanding the dynamic of atmospheric pollution is an important tool for managing air quality. The use of statistical learning tools has not been widely used in the study of air pollutants. The purpose of this work is to identify the key variables controlling daily and seasonal variations of PMcoarse, PM2.5 and CO concentrations. This analysis is performed to untangle the contribution from local and regional or meteorological factors impacting air quality. Two statistical learning techniques were used: Classification and Regression Trees (CART) and Random Forest- The latter combines up to 1000 classification trees with a majority-vote weighting criterion. Wind speed, wind direction, temperature, precipitation, humidity, solar radiation and barometric pressure were used as local meteorological input variables. Daily Aerosol Optical Depth (AOD) segregated into BC, OM, Dust, sulfate from CAMS-Copernicus were retrieved and used as local and regional chemical input variables. A long-record of radiosonde profiles associated to boundary layer structure also were used as local meteorological input influencing variations of pollutant levels. Back-trajectories of air masses arriving Bogotá were performed to analyze the long-range transport of pollutants. The number of MODIS active fires and their fire radiative (FRP) power from Northern South America selected in the vicinity of the air masses arriving Bogotá were included as regional variables. All analyses were performed daily for the period ranging from January 2008 to December 2018. Classification Tree and Random Forest techniques were implemented through R script to classify PMcoarse, PM2.5, and CO concentrations into 5 categories (5 quantiles), finding the best variables deciphering changes between concentration categories. The main variable that better classified seasonal concentration levels of pollutants was the Fire Radiative Power, and the local pressure in Bogotá as indicator of the intertropical Convergence Zone. Anomalous levels of concentration were described mainly by mixing height, wind direction, and wind speed. Moreover,  AOD segregated was important to understand the seasonal and daily variations.