Braluft: Forecasting air quality using incremental models and computer vision

Abstract

Air quality in urban areas is an issue of great concern as it affects public health and local environments. By forecasting the pollutant levels public administrations may be notified of periods with potentially bad air quality and can initiate strategic policies to limit the spreading of pollutants. One of the challenges associated with forecasting air quality is the fact that meteorological conditions and anthropogenic activities change as seasons passes. This thesis targets such issues by presenting Braluft, a distributed system designed to incrementally train forecasting models over time using machine learning. The thesis makes use of the program to evaluate: (a) which variables influence the levels of two important pollutants, NO2 and PM10, at Danmarksplass, Bergen, and (b) whether the incremental approach is well suited for making air quality forecasts by continuously adjusting to new observations. The program uses weather forecasts and traffic level as input data, and the latter is assessed by applying computer vision to a web camera overlooking the area. The most promising variables for NO2 forecasting turned out to be wind speed and traffic levels by a wide margin. PM10 levels are seemingly a result of more complex processes where all the observed variables have an influence. The program delivers promising results for its intended purposes, namely register trends occurring in the air quality and subsequently make air quality forecasts based on these trends. This results in good air quality forecasts for most days where the pollutant levels are low. However, bad air quality is often a result of sudden changes and can hardly be considered a trend. The program is therefore struggling to foresee such events. The concept supporting the program might prove more valuable in areas where raises in pollutant levels are less abrupt.