Contextualized spatio-temporal graph-based method for forecasting sparse geospatial sensor networks

Spatio-temporal forecasting is a rapidly evolving field, accelerated by the increasing accessibility of sensoring infrastructure and computational hardware, capable of processing the large amount of sampled data. Applications of spatio-temporal forecasts range from traffic, weather, air pollution forecasting and others. Emerging technologies employ deep learning architectures, such as graph, convolutional, recurrent and transformer neural networks. While the state-of-the-art methods provide accurate time series predictions, they are typically limited to providing forecasts only for the direct locations of sampling, whereas coverage of the entire area is often desired by the applications. In this work, we propose a method that addresses this challenge and improves on the shortcomings of related works, which have already tackled the task. The proposed graph convolutional recurrent neural network based method provides forecasts for arbitrary geolocations without available measurement data, formulating predictions based on contextual information of target geolocations and the time series data of nearby measurement geolocations. We evaluate the method on three real-world datasets from meteorological, traffic and air pollution domains, and gauge its performance against the state-of-the-art spatio-temporal forecasting methods. The proposed method achieves 12.26 %, 66.97 % and 42.89 % improvements in the mean absolute percentage error