Geospatial Timeseries Imputation using Deep Neural Networks

Abstract: With the advancement of technology, data collection has become a big part of many industries.Large amounts of data can be used for analytics purposes, and give companies the opportunityto offer a wider range of services to their customers. Due to the ever-increasing interest in data,and due to restrictive regulations such as the General Data Protection Regulation (GDPR),synthetic data generation is gaining attention from researchers in both industry and academia. Scania collects vehicle data from more than 450,000 trucks and buses worldwide, and use thisfor diagnostics and fleet management purposes. However, the frequency of data collection istailored to customer needs, and Scania could still benefit from having access to higher temporalresolution data. This project investigates the possibility of using deep learning methods for the purpose ofincreasing the temporal resolution in geospatial time series. Specifically, a classic feed forwardnetwork as well as a Generative Adversarial Network (GAN) is evaluated and compared to asimpler, more intuitive nearest neighbour based baseline model. While deep neural network complexity allow for reasonable quantitative results, qualitativeevaluations show that the spatial context of geospatial data is hard for the models to learn.Without any conditioning on road networks, the models generate vehicle positions that are notalways realistic.