Resource efficient data transmission on animals based on machine learning

TL;DR

This study demonstrates that on-board pattern recognition using lightweight decision trees can selectively transmit biologically relevant data from bio-loggers, significantly reducing energy consumption and extending device runtimes while maintaining meaningful behaviour detection. By training trees off-board on PC and deploying compact classifiers on the WildFi tag, the approach achieves high accuracy for specific behaviours and can cut transmission data by substantial margins (e.g., ~14.68{\%} in a standing-detection scenario, up to near 100{\%} in extreme cases). The gyroscope emerges as a crucial feature for distinguishing postural behaviours, and energy savings scale with the proportion of data filtered or selectively transmitted. Overall, the work provides a practical pathway to longer-term, energy-efficient wildlife monitoring through hardware-agnostic, software-driven data management.

Abstract

Bio-loggers, electronic devices used to track animal behaviour through various sensors, have become essential in wildlife research. Despite continuous improvements in their capabilities, bio-loggers still face significant limitations in storage, processing, and data transmission due to the constraints of size and weight, which are necessary to avoid disturbing the animals. This study aims to explore how selective data transmission, guided by machine learning, can reduce the energy consumption of bio-loggers, thereby extending their operational lifespan without requiring hardware modifications.

Figures (5)

• Figure 1: An illustration of the WildFi tag from the original publication Wild2022. On the left is a coin to give a perspective on the device's size. In the centre is the main device from front and, on the right, from back.
• Figure 2: This image illustrates the concepts behind the reduction of transmission time with the help of contextual transmission. "Regular" means to send every message whole. "Selected" means to send only a selection of information. "Conditional" means to send messages when a condition is met. "Both" corresponds to a combination of conditional and selected data transmission. Each blue box indicates a message that is being sent. The dotted lines illustrate that a portion of a message is not transmitted. The bottom arc shows a progression over time, without a specified unit of time, to allow for an abstract comparability between the different approaches.
• Figure 3: This is a modified version of the activity recognition chain pipeline by Bulling et al. bulling2014tutorial. From left to right, the single boxes show the abstract steps that are appropriate to achieve a successful classification. We additionally marked which step happens on which type of device, either PC or bio-logger, by colouring the PC steps grey and the onboard steps blue. The original author permitted us to use their graphic.
• Figure 4: This plot compares the actual behaviours with the classified behaviours throughout the experimental data. The y-axis indicates the different behaviours. The x-axis shows the timesteps. The blue line shows which behaviour was exhibited during the recording, therefore mapping to the actual labels of the time series data. The grey lines connect the behaviour classified by the decision tree with the actual behaviour at that particular timestep. Therefore, the grey lines indicate wrong classifications.
• Figure 5: This image shows the confusion matrix for the decision tree, optimised for the behaviour standing, with a tree depth of 7, belonging to Participant 1. The decision tree has an increased number of misclassifications between sitting, standing and walking. The y-axis shows the actual behaviours and the x-axis what the classifier has predicted.