Raw accelerometer data and annotated video recordings of behaviours during a nutritional challenge: a dataset for training behaviour classification models in goats

Raw accelerometer data: Fourteen Alpine dairy goats were equipped with accelerometers attached to collars. The accelerometers were programmed to record the acceleration on the x-y-z-axis at a frequency of 10 Hz for 19 consecutive days with a 2-day nutritional challenge (raw_data files). The experiment consisted in a seven-day control pre-challenge period, with a standard lactation TMR (Total Mixed Ration), two days of straw feeding only (challenge period), and a post-challenge period of ten days on the standard TMR. The challenge period occurred from 2023-09-29 15:00:00 to 2023-10-01 15:00:00. Raw_data files are .txt files with 5 columns; « date » is the date and time of the beginning of the recording in the format « YYYYMMDDHHmmss », « t » is the amount of seconds since the beginning of the recording, and « x », « y » and « z » are the acceleration values recorded on the three axis. training_data files: A total of 180 hours of annotated data with acceleration values from five of the fourteen goats previously placed in the same experimental setup is provided in this dataset. “Ruminating”, “head in the feeder”, and “lying” behaviours were annotated from video recordings using the Boris software for 180 hours by a single trained observer using a pre-established ethogram. Training_data files are .csv files with 6 columns; « TIME » is the date and time, « ACCx », « ACCy » and « ACCz » are the acceleration values recorded on the three axis, « feeding_behav_data_goat » is the feeding behaviour annotated from video recording according to the ethogram (provided), « position_behav_data_goat » is the positional behaviours annotated from video recording according to the ethogram (provided). Result file: The training data files were used in the ACT4Behav pipeline (Mauny et al., 2024) to train three models able to classify accelerometer data into “ruminating”, “head in the feeder”, and “lying” behaviours. These models were then applied to the raw accelerometer data using 60-second time windows. Features were computed from these windows, and the outputs were compiled into the results_data_features.csv file. The first column of this file is id_goat, the second is time_window, which indicates the date and time corresponding to the beginning of each 60-second window, and the remaining columns contain the features described in the following preprint: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5360946.