Deep learning accurately predicts white shark locomotor activity from depth data

Deep learning accurately predicts white shark locomotor activity from depth data

Zac Yung-Chun Liu, Jerry H. Moxley, Paul KaniveAdrian C. Gleiss, Thom Maughan, Larry Bird, Oliver J. D. Jewell, Taylor K. Chapple, Tyler Gagne, Connor F. White, Salvador J. Jorgensen

ABSTRACT:

Background

The study of bioenergetics, kinematics, and behavior in free-ranging animals has been transformed through the increasing use of biologging devices that sample motion intensively with high-resolution sensors. Overall dynamic body acceleration (ODBA) derived from biologging tags has been validated as a proxy of locomotor energy expenditure has been calibrated in a range of terrestrial and aquatic taxa. The increased temporal resolution required to discern fine-scale processes and infer energetic expenditure, however, is associated with increased power and memory requirements, as well as the logistical challenges of recovering data from archival instruments. This limits the duration and spatial extent of studies, potentially excluding relevant ecological processes that occur over larger scales.

Method

Here, we present a procedure that uses deep learning to estimate locomotor activity solely from vertical movement patterns. We trained artificial neural networks (ANNs) to predict ODBA from univariate depth (pressure) data from two free-swimming white sharks (Carcharodon carcharias).

Results

Following 1 h of training data from an individual shark, ANN enabled robust predictions of ODBA from 1 Hz pressure sensor data at multiple temporal scales. These predictions consistently out-performed a null central-tendency model and generalized predictions more accurately than other machine learning techniques tested. The ANN prediction accuracy of ODBA integrated overtime periods ≥ 10 min was consistently high (~ 90% accuracy, > 10% improvement over null) for the same shark and equivalently generalizable across individuals (> 75% accuracy). Instantaneous ODBA estimates were more variable (R² = 0.54 for shark 1, 0.24 for shark 2). Prediction accuracy was insensitive to the volume of training data, no observable gains were achieved in predicting 6 h of test data beyond 1–3 h of training.

Conclusions

Augmenting simple depth metrics with energetic and kinematic information from comparatively short-lived, high-resolution datasets greatly expands the potential inference that can be drawn from more common and widely deployed time-depth recorder (TDR) datasets. Future research efforts will focus on building a broadly generalized model that leverages archives of full motion sensor biologging data sets with the greatest number of individuals encompassing diverse habitats, behaviors, and attachment methods.

Animal Biotelemetry (2019) 7: 14, DOI: 10.1186/s40317-019-0175-5

SOURCE