Abstract
Monitoring air pollution is of vital importance to the overall health of thepopulation. Unfortunately, devices that can measure air quality can beexpensive, and many cities in low and middle-income countries have to rely on asparse allocation of them. In this paper, we investigate the use of GaussianProcesses for both nowcasting the current air-pollution in places where thereare no sensors and forecasting the air-pollution in the future at the sensorlocations. In particular, we focus on the city of Kampala in Uganda, using datafrom AirQo's network of sensors. We demonstrate the advantage of removingoutliers, compare different kernel functions and additional inputs. We alsocompare two sparse approximations to allow for the large amounts of temporaldata in the dataset.