Can advanced machine learning techniques help to reconstruct barn swallows’ long-distance migratory paths?

Introduction Light level geolocators are a well-established technology to track migratory animals (like passerine birds) that are too small to carry satellite tags. Recently, different methods to reconstruct migration paths from geolocator data have been developed (e.g. FLightR R-package). All of these require preprocessing of raw data: light measurements are often affected by differential shading, and twilight events showing unnatural variation in light levels and timing need to be corrected (Fig. 1). Here we propose and implement advanced machine learning (ML) techniques to automate this procedure (Fig. 2) and compare performance to that of expert human editors. Methods We rely on nearly 37,000 expert-classified twilight events from almost a hundred geolocators applied to a longdistance migratory bird: the barn swallow (Hirundo rustica). We selected as predictors 8 intra-twilight luminosity measurements (first/last 8 light measurements after sunrise/before sunset), 18 inter-twilight values (both twilight times of each night on a moving window of 9 days), and 4 expert-defined aggregated statistics of these. Implementing a Logistic Regression (LR) classifier as benchmark, we next considered advanced machine learning models such as a Random Forest (RF) and a deep Neural Network (NN). We investigated different architectures such as the effect of adding to the classical fully connected structure some 1D convolutional layers, which usually turn out to be efficient in dealing with time series. Results Our analysis shows that ML can help to correctly filter geolocator data. The complex models here proposed outperform the LR considering both overall accuracy (LR: 77.91%, RF: 89.15%, NN: 89.40%) and confusion matrix as metrics (Fig. 3). Comparison between migratory tracks of a test individual estimated by FLightR using twilights classified according to different methods shows that ML algorithms can delete noisy twilight events that could significantly distort the result (Fig. 4). Discussion In this work we showed how ML algorithms can perform better than classical statistical learning in a complex classification task. Migratory paths obtained after preprocessing with the structures proposed are closely similar to the ones computed after manual editing by an expert. This work could constitute the missing part to the complete automation of geolocator data analysis.