Signal selection in a complex environmental distributed sensing problem

Supporting sustainable development for the urban environment is crucial in the age of rapid urbanisation. Air pollution modelling is one of the key tools for researchers, scientists, and urban planners to understand pollution behaviour. Recent updates in air quality regulations are challenging the state-of-the-art air pollution modelling techniques by requiring accurate predictions on a high temporal level, i.e. predictions at the hourly level rather than the annual level. Current state-of-the-art models designed to have good prediction accuracy on the low temporal resolution by assuming that the pollution is in steady state. Making predictions on higher temporal resolution violates this assumption and causing inaccurate predictions. We introduce a novel statistical regression based air pollution model which produces accurate hourly predictions by using data with high temporal resolution and advanced regression algorithms. We conducted an analysis which shows that the state-of-the-art evaluation techniques (e.g. RMSE) do not describe the nature of the mispredictions of the models built on different data subsets. We carried out an extensive input data evaluation experiment where we concluded that our approach could achieve further accuracy improvement by training the models on a carefully selected subset of the input data.