Calibration of Spatial Forecasts from Citizen Science Urban Air Pollution Data
Explore a machine learning approach for forecasting high-frequency spatial fields of urban air pollution using citizen science data in this 56-minute talk by Stefano Castruccio at the Alan Turing Institute. Learn about a novel method that incorporates sparse recurrent neural networks with a spike-and-slab prior, offering advantages over standard neural network techniques. Discover how this approach achieves a smaller parametric space and introduces stochastic elements to generate additional uncertainty. Understand the importance of forecast calibration, both marginally and spatially, and its application to assessing exposure to urban air pollution in San Francisco. Gain insights into the significant improvements in mean squared error compared to standard time series approaches, with calibrated forecasts extending up to 5 days.