Using non-continuous accelerometry to identify cryptic nesting events of Galapagos giant tortoises

Triaxial accelerometers have revolutionized wildlife research by providing an unprecedented understanding of the behavior of free-living animals. Machine learning is often applied to acceleration data to classify diverse animal behaviors across taxa. However, the high frequency, continuous data collection typically favored for behavioral classification studies often generates very large data sets, which may inhibit remote data acquisition and make data storage challenging. Coarse-frequency sampling or non-continuous bursts of acceleration data reduce these problems. To analyze such data, a suite of variables that summarize key features of the behavior of interest can be generated. These variables can then be used in numerous classification approaches, accommodating variation in data collection methods or sampling regimes. We demonstrate the potential for non-continuous accelerometer data to identify long-duration behavior and employ machine learning to classify the nesting behaviors of the critically endangered eastern Santa Cruz giant tortoise (Chelonoidis donfaustoi). We field validated 112 nesting events from 21 giant tortoises. We then derived summary statistics based on accelerometry (e.g., overall dynamic body acceleration, metrics comparing acceleration before and after the probable event) and used them as inputs for Random Forest and Boosted Regression Tree classification algorithms. Our models produced a harmonic mean of precision and sensitivity (F1-score) of 0.91. We tested the generality of our model and found that the model performs well when applied to both novel individuals and years. The most important variable in accurately classifying data sequences was the proportion of acceleration data bursts above an activity threshold followed by the average overall dynamic body acceleration value of the bursts. These results demonstrate the feasibility and efficacy of using non-continuous accelerometer data to identify prolonged, biologically relevant behaviors in free-living wildlife. By using summary variables that do not require continuous sampling, this approach facilitates long-term monitoring of animal behavior. Similar methodology has potential to inform priority questions in ecology and conservation, such as predicting wildlife responses to climate change and identifying critical habitats, with applications across diverse species and behaviors.