Accelerometer, gyroscope and pressure data associated with behaviors of free-ranging hawksbill sea turtles (Eretmochelys imbricata)

In this paper, we explored the use of transfer learning across species and taxa, employing fully convolutional neural networks to predict the behaviors of critically endangered hawksbill sea turtles from acceleration data. For this purpose, fully convolutional neural networks (V-net and U-net) were pre-trained on a dataset of green turtles (Zenodo link) and human data (Intensive Care Unit (ICU) HAR dataset, https://doi.org/10.24432/C54S4K) before being fine-tuned on the hawksbill dataset. The results reveal a 8% and 4% improvement in F1-score with transfer learning from the green turtle and human datasets, respectively, compared to training the models from random weight initialization (without transfer learning). 

The dataset comprised the raw acceleration, gyroscope and depth sequence of 6 free-ranging hawksbill sea turtles associated with the behaviors. The indiviuals were equipped with a on-board video recorder combined with an accelerometer, gyroscope, magnetometer and luminosity, temperature and depth sensors using four suction cups and an automatic release system over a two-day periods (see Jeantet et al. 2020 for details and the associated article). The accelerometer, gyroscope, magnetometer recorded at 20 Hz and the pressure, temperature and luminosity sensors at 1 Hz. The cameras were programmed to record until nightfall (6 pm) and resume at daybreak (6 am). The magnetometer, luminosity and temperature data are provided but not used in the associated study. 

For each individual, the data collected by the devices was correlated with observed behaviors from video recordings. Unlabeled sequences, mostly comprising night recordings, were excluded, resulting in the creation of one file per day of deployment for each individual. 

To process the depth data and increase the sampling rate to 20 Hz, we used a linear interpolation technique. We called this new variable "Pressure_corr". Additionally, we calculated the pressure difference ("Pressure_diff") between each measuring point (originally at 1 Hz).

"In total, 69.7 hours of multi-sensor sequences were labelled from six different hawksbill turtles (approximately 11.6 hours of recording per individual, max = 17.8 hours, min = 6.3 hours, standard deviation = 3.6 hours). The predominant behavior observed in the videos was Feeding, totaling over 38.6 hours, followed by Resting and Swimming, with 19.1 hours and 7.9 hours, respectively. The other behaviors were expressed in minority (Breathing: 2.2 hours, Gliding: 1 hour, Scratching: 0.8 hour and Other: 0.1 hour). " 

The folder contains 10 Python matrices, each with 15 columns (AccX, AccY, AccZ, GyrX, GyrY, GyrZ, MagX, MagY, MagZ, Depth, Light, Temperature, Pressure_corr, Pressur_diff, Behavior) and a number of rows corresponding to the deployment duration. The title of each file indicates the camera number used (CC-09-XX) and the deployment day (DD-MM-YYYY), with the last digit specifying whether the matrix corresponds to the first or second day of deployment.

The folder also contains two dictionaries (behInd_to_behName, behName_to_behInd) that specify the behaviors associated with each number used as a label in the Behavior column. Additionally, there is a dictionary (dico_info) that provides the names of the matrix columns and the frequence of recording.