Reconstruction of long-distance bird migration routes using advanced machine learning techniques on geolocator data.
Geolocators are a well-established technology to reconstruct migration routes of animals that are too small to carry satellite tags (e.g. passerine birds). These devices record environmental light-level data that enable the reconstruction of daily positions from the time of twilight. However, all current methods for analysing geolocator data require manual pre-processing of raw records to eliminate twilight events showing unnatural variation in light levels, a step that is time-consuming and must be accomplished by a trained expert. Here, we propose and implement advanced machine learning techniques to automate this procedure and we apply them to 108 migration tracks of barn swallows ( Hirundo rustica). We show that routes reconstructed from the automated pre-processing are comparable to those obtained from manualselection accomplished by a human expert. This raises the possibility of fully automating light-level geolocator data analysis and possibly analysing the large amount of data already collected on several species.
deep neural network; light-level tag; migratory species; movement ecology; path estimation; random forest