14Animal-borne data loggers, i.e., biologgers, allow researchers to record a variety of sensor data 15 from animals in their natural environments (Hussey et al. 2015; Kays et al. 2015). This data allows 16 biologists to observe many aspects of the animals' lives, including their behavior, physiology, social 17 interactions, and external environment. However, the need to limit the size of these devices to a 18 small fraction of the animal's size imposes strict limits on the devices' hardware and battery 19 capacities (Kays et al. 2015). Here we show how AI can be leveraged on board these devices to 20 intelligently control their activation of costly sensors, e.g., video cameras, allowing them to make 21 the most of their limited resources during long deployment periods. Our method goes beyond 22 Sensor data logger
104We begin with a brief introduction to the sensor data loggers used in this study (for more details see 105 Online Methods). Fig. 1 (c) shows a close-up view of the logger, with the camera module located on 106 the far-left end of the logger. Fig. 1 (d) shows an example of the data collected from a chest-107 mounted logger, with the map displaying the GPS data collected and the two inset images showing 108 frames from foraging activity captured by the device. Fig. 1 (e) shows an example of how these 109 devices were attached in the field. In this example, the logger is attached on the back of the animal, 110 with the camera facing forward and the GPS receiver (white square to the rear of the device) facing 111 the sky. Additionally, in some cases the devices were instead attached to the chest of the birds, in 112 order to improve the camera's field of view during foraging activities.
114Note that because our logger is equipped with a commercially-available MCU and sensors using a 115 simple circuit design (see Online Methods), we believe that reproduction of the logger system using 116 rapid prototyping platforms, such as Arduino, is relatively easy.
118Activity recognition method 119 Overview 120 Our method is based on supervised machine learning, which can be divided into two main phases: 121 training and testing. This approach assumes that sensor data that corresponds to the data collected 122 by our low-energy sensors can be collected in advance during the training phase. During the 123 training phase, the preexisting sensor data is labelled by biologists to indicate the target activities 124 that should be captured by the loggers' cameras. This labeled sensor data is then used to train the 125