Spatial Interpolation of Air Quality Data with Multidimensional Gaussian Processes

Abstract

The central question of this paper is whether interpolation techniques applied to a distributed sensor network can indeed provide more information than using the constant background of an urban reference station to measure air pollution. We compare different interpolation techniques based on temporal-spatial machine learning in terms of their applicability for correctly predicting personal exposure. Using a dataset of stationary low-cost sensors, we estimate exposure on a route through the city and compare it to mobile measurements. The results show that while different machine learning-based interpolation methods yield quite different results, validation of machine learning-based approaches is still challenging.