Predicting the fine-scale spatial distribution of zoonotic reservoirs using computer vision

Abstract

Zoonotic diseases threaten human health worldwide and are often associated with anthropogenic disturbance. Predicting how disturbance influences spillover risk is critical for effective disease intervention but difficult to achieve at fine spatial scales. Here, we develop a method that learns the spatial distribution of a reservoir species from aerial imagery. Our approach uses neural networks to extract features of known or hypothesized importance from images. The spatial distribution of these features is then summarized and linked to spatially explicit reservoir presence/absence data using boosted regression trees. We demonstrate the utility of our method by applying it to the reservoir of Lassa virus, Mastomys natalensis, within the West African nations of Sierra Leone and Guinea. We show that, when trained using reservoir trapping data and publicly available aerial imagery, our framework learns relationships between environmental features and reservoir occurrence and accurately ranks areas according to the likelihood of reservoir presence.