An RSSI-based localisation method for fine-scale wildlife tracking using an Automated Radio Telemetry System

CHAPTER FOUR An RSSI-based localisation method for fine-scale wildlife tracking using an Automated Radio Telemetry System

Abstract

4.1 Introduction

Movement is fundamental to every aspect of life on earth. It is universal among organisms, shapes individual fitness, directs evolutionary development, and influences ecological processes, including responses to anthropogenic change. Understanding how, where, when and why animals move is therefore essential to answer questions in ecology, animal behaviour and conservation sciences. Tracking wild animals has revealed detailed patterns of habitat use; discovered long-term migrations across the globe; uncovered mechanisms of navigation, memory and cognition; and characterised animal communication and social behaviour. Despite this importance, the study of animal movement has proved challenging in many taxa, due to the difficulty associated with collecting individual-level data from wild animals to an accurate degree.

Recent advances in wildlife tracking technologies, data processing tools and analytical techniques has allowed researchers to monitor many different terrestrial and aquatic organisms in almost continuous space and time, resulting in the growth of the field of movement ecology. Sensors that are widely used to track wild animals include Global Positioning Systems tags, proximity loggers, passive integrated transponder tags, very high frequency or ultra high frequency radio tags, wireless sensor networks and bio-acoustic microphone arrays. Each method comes with major strengths and weaknesses, meaning there is no one-perfect solution. In general, the main trade-off's of tracking involve tag mass, sampling frequency, data retrieval options, cost, accuracy, and battery life. For example, passive integrated transponder tags utilise radio-frequency identification detection antennas which detect changes in the electromagnetic field. Although these tags are cheap and lightweight, they require the animal to come into close proximity of approximately ten centimeters of the antenna, which limits the ecological systems that this method can be applied to and incurs high risk of sampling biases. Bluetooth low-energy beacons are another low cost alternative, which utilise Apple's Find My network to locate tags. Although high-levels of accuracy can be obtained in a phone-rich urban environment, this strong sampling bias which affects both the spatial and temporal likelihood of tag localisation makes implementing this method unfeasible for animals in wild and remote spaces.

Tags which utilise Global Positioning Systems remain the primary choice for animal tracking due to their high spatio-temporal accuracy and global application. However due to the high energy demand needed for accurate localisation, Global Positioning Systems tags require heavy batteries and therefore remain too large to track around sixty-five percent of the worlds mammal species and seventy percent of birds. The launch of the international cooperation for animal research using space initiative is a promising move in the continued miniaturization of Global Positioning Systems tags and worldwide tracking.

et al. 2022). Although in the future Global Positioning Systems tags may be reduced to an appropriate mass for small vertebrates, the limitations of Global Positioning Systems signals in complex habitats such as trees, caves or underground burrows make this option infeasible for animals which utilise these spaces. Other solutions with competing spatio-temporal resolution have become available, such as the Wildlife Biologging Network and the ATLAS system, but cost and installation effort remain high and technical expertise are required to operate the systems.

Radio transmitters have been used to monitor animal movement since the early nineteen sixties. Today they remain a popular, and often the only, option for tracking small animals due to the low-cost and low-weight of the tags. Traditional radio tracking involves locating and concurrently triangulating individually tagged animals in the landscape by trained field ecologists using hand-held receivers. This method is both logistically challenging and labour intensive and often the resulting data is irregularly sampled, of low spatial-temporal accuracy, and limited to few individuals. Further to this, there is often no attempt to evaluate the efficacy of traditional radio tracking when utilised in the field and there has been little work to evaluate the implemented techniques of positional error estimation, which is a requirement for applying many spatial models.

A major leap forward in the use of radio telemetry for wildlife tracking has been the development of Automated Radio Telemetry Systems, which have enabled often hundreds of animals to be simultaneously and continuously tracked across large areas in a range of habitats. Different Automated Radio Telemetry Systems designs have been implemented, but in general the main framework consists of a network of static radio receivers which localise the transmitter via indirect measures of the received signal. Automated Radio Telemetry Systems are generally split into omni-directional and directional systems, which vary in their receiver type, detection capabilities and cost. Omni-direction systems typically have simpler receivers which use one isotropic antenna to detect in an approximately uniform pattern. Directional systems utilise more complex receivers which have multiple Yagi-Uda antennas that are each set in a different orientation. The different Automated Radio Telemetry Systems designs come with their own drawback, the most notable being the complexity of installation, optimization and operation; and additionally the difficulty associated with estimating transmitter locations from received signal strength indicator.

The relationship between tag signal strength and the distance to the receiver can theoretically be described as an exponential decay function. In practice, the signal attenuation from environmental conditions (e.g., precipitation, habitat type) and multipathing effects, such as shadowing (signal blocked by object) and reflection (signal bounce), adds significant noise to RSSI making the true relationship difficult to predict in the environment. A diverse number of range-based localisation methods have been developed from this RSSI-Distance relationship, which includes lateration techniques. Lateration is commonly applied to omni-directional system designs and requires simultaneous detections from at least three receivers to estimate a location using the receivers overlap. Directional ARTS typically apply direction-based techniques for localisation, such as angulation estimation using the angle of arrival. This method estimates the angle of arrival of the transmitter to the receiver by comparing the relative RSSI between antennas of the same receiver. Angles of arrival from at least two receivers are then used in triangulation. Due to the requirements of multiple receiver detections, location drop out rate is often high for both lateration and angulation techniques. Positional error remains substantially higher than GPS-based systems, but techniques to minimise error have been implemented, including the use of RSSI-filters, grid search algorithms and trajectory smoothing with Kalman Filter.

Location fingerprinting is a range-free localisation technique that has been widely implemented in indoor positioning systems. The basic principal of this method is to develop a site-specific radio map of known reference points to train a machine learning model, which is then used to estimate the locations of unknown signals. The radio map characterises the RSSI at known locations, which encompasses the effect of the surrounding environment on signal noise. For outdoor application, promising work has been conducted in the recent emergence of Internet of Things technologies, which is connecting everyday objects and device in a large low-powered communication network. Studies using an open-source dataset from the Sigfox Internet of Things has shown fingerprinting to be a highly accurate method when receiver coverage is high. The Sigfox Internet of Things has been used in animal tracking however accuracy remains coarse in remote spaces (one to ten kilometers resolution), tag weights remain too large for small vertebrates (greater than one gram) and data coverage is mostly restricted to countries in Europe. Both Wallace et al. and Tyson et al., have successfully implemented location fingerprinting methods for an omni-directional system, which were appropriate for tracking small vertebrates in a compact study area. Collecting the suitable training datasets for these systems can be extremely labour intensive and requires constant updating when the network setup is altered. As shown by Osta et al., an alternative approach is to implement a general fingerprinting model to receivers of the same design and in a similar environment. This approach comes at a cost of overall accuracy and has to date only been implemented in a directional system.

In this study, we present an open-source localisation method to estimate radio transmitter positions from an automated radio telemetry system. We have developed two machine learning pipelines which consist of a shared general receiver model that can pool training data across similar receivers, and an extended receiver model which utilises the concept of fingerprinting to improve accuracy. We demonstrate the utility of this methodology using data collected from an ARTS with multiple receiver types in an approximately four thousand five hundred hectares region of Devon, South-West England, which is tracking a bat species of conservation concern in Britain, the greater horseshoe bat. Using a series of controlled transect, training and test datasets were created to compare the performance of our general and extended receiver models with two other published fingerprinting-based methods: fingerprinting and receiver offsetting. We additionally evaluated the systematic error in the ARTS and utilised the resulting models to predict the associated error of an unknown location. Finally, we developed a rule-based filtering approach to improve positional error and assessed the effect on mean and median error and location drop out rate.