Modelling Satellite Data for Automobile Insurance Risk

High-resolution satellite data opens new applications for personalised hyper-local insurance risk scores. In this paper, open-source satellite imagery and accident incident data are aggregated to model road network risk for the generation of a risk map to be used in insurance quotation. Two machine learning approaches are outlined - an unsupervised clustering of the images based on extracted features and a supervised labelled data approach. These are tried against the data; results illustrating the potential utility of this approach for estimating insurance risk are given. For the unsupervised approach, distinct risk levels were produced which reflected historic levels of accident for visually similar road sections. With supervised methods, we performed a binary classification to identify accident sites with a recall of 82% and precision of 80%.