Machine learning approach to forecasting urban pollution

Abstract

This work addresses the question of how to predict fine particulate matter given a combination of weather conditions. A compilation of several years of meteorological data in the city of Quito, Ecuador, are used to build models using a machine learning approach. The study presents a decision tree algorithm that learns to classify the concentrations of fine aerosols, into two categories (>15μg/m3 vs. <;15μg/m3), from a limited number of parameters such as the level of precipitation and the wind speed and direction. Requiring few rules, the resulting models are able to infer the concentration outcome with significant accuracy. This fundamental research intends to be a preliminary step in the development of a web-based platform and smartphone app to alert the inhabitants of Ecuador's capital about the risk to human health, with potential future application in other urban areas.