Classification of birds and bats using flight tracks

Highlights

• Modeled classification is needed for assessing collision risk from wind turbines.

• We classified birds and bats from flight tracks from a single-camera thermal video.

• We classified terns, swallow, bats, and gulls with an average 82% accuracy.

• Measures of distance and wingbeat frequency would improve model.

• A larger annotated library of flight tracks from thermal video is needed.

Abstract

Classification of birds and bats that use areas targeted for offshore wind farm development is essential to evaluating the potential effects of development. The current approach to assessing the number and distribution of birds at sea is transect-surveys conducted by trained individuals in boats or planes, or analysis of imagery collected from aerial surveys. These methods can be costly and pose safety concerns so that observation times are limited to daylight hours and fair weather. We propose an alternative method based on analysis of thermal video that could be recorded autonomously. We present a framework for building models to classify birds and bats and their associated behaviors from their flight tracks. As an example, we developed a discriminant model for theoretical flight paths and applied it to data (N = 64 tracks) extracted from 5-minute video clips. The agreement between model- and observer-classified path types was initially only 41%, but it increased to 73% when small-scale jitter was censored and the number of different path types was reduced. Classification of 46 tracks of bats, swallows, gulls, and terns on average was 82% accurate, based on a jackknife cross-validation. Model classification of gulls and swallows (N ≥ 18) was on average 73% and 85% correct, respectively. Model classification of bats and terns (N = 4 and 2, respectively) was 94% and 91% correct, respectively; however, the variance associated with the tracks from these targets is poorly estimated. The models developed here should be considered preliminary because they are based on a small data set both in terms of the numbers of species and the identified flight tracks. Future classification models could be improved if the distance between the camera and the target was known.

Introduction

The development of renewable energy sources, including wind, has become part of the future U.S. energy portfolio in an effort to reduce the dependence and the environmental impacts associated with extracting, transporting, and burning fossil fuels. In order to build and operate a wind farm, developers must abide by Federal and State conservation laws including the National Environmental Policy Act, Migratory Bird Treaty Act, and the Endangered Species Act. Each law has its own unique requirements, so the U.S. Fish and Wildlife Service (USFWS) provided guidelines to ease the process for developers of land-based wind energy (USFWS, 2012). These guidelines rely upon pre- and post-construction surveys using field personnel to determine the presence and abundance of birds in order to model and validate the risk to birds. As the annual production of electricity from land-based wind farms has grown, so has the interest in development of production-scale offshore wind energy. However, empirical data describing the distribution, abundance, behavior, and life history is lacking for many bird species. Further, land-based methods to gather these data are either impossible or prohibitively expensive offshore.

European offshore wind energy development precedes the U.S. and already includes 69 wind farms in eleven countries (EWEA, 2014). Methods used to assess risk and effects to bird populations at multiple scales during pre- and post-construction included direct observation or video surveys along boat and aerial transects (Banks et al., 2005, Camphuysen et al., 2004), collision risk models (Band, 2012, Smales et al., 2013), and field experiments to assess turbine avoidance (Guillemette and Larsen, 2002). As research efforts in Europe have gained a greater understanding of effects from wind farms, the need for additional tools and techniques to appropriately assess impacts has become apparent (Bailey et al., 2014). Efforts have been initiated to gather broad-scale wildlife distribution and abundance information off the Atlantic and Pacific coasts of the U.S. to assess potential offshore wind-wildlife conflicts and to aid in the leasing of offshore sites for wind energy development (Adams et al., 2014, Maclean et al., 2009, Normandeau Associates, Inc. (Normandeau), 2012, U.S. Geological Survey (USGS), 2014). Research projects are also being conducted to gather empirical data necessary to model risk to birds from collision mortality with wind turbines operating offshore (BOEM, 2014). The limitations of using field personnel for offshore surveys will likely result in an increased use of remote sensing technologies to gather the data necessary to validate these risk models.

Although remote-sensing tools are already used to characterize offshore wildlife populations, expenses related to data interpretation are still very high and automated detection and classification methods are not yet well developed. For instance, Normandeau (2012) used automated detection algorithms on high spatial resolution imagery gathered for broad-scale assessment to screen out image frames without targets of interest and manually-reviewed the remaining images to glean data. Although Buckland et al. (2012) did not include cost as a factor in comparing digital and visual aerial survey methods, the authors did note that locating and identifying birds in images could be automated from digital survey imagery. Similarly Groom et al. (2013) used large format digital aerial cameras to record aerial surveys for marine birds. Image processing was used to extract regions within images that contained bird-like objects, and these data were then examined by trained ornithologists. Because of the intense effort involved in manual data review, an automated animal-detection algorithm that is both cost-effective and scientifically valid was deemed essential for offshore risk assessment (Normandeau, 2012).

Radar has been used to survey wildlife both day and night for site scale assessments of wind farm collision risk and avoidance behavior. Numerous types of radar have been used to detect, track, identify and study bird behavior including high-resolution marine surveillance radar near offshore wind farms (Gauthreaux and Belser, 2003, Plonczkier and Simms, 2012). Target periodic amplitude modulation of a radar echo has been used as an indicator of wing beat frequency and thus used to identify birds (Bruderer et al., 2010). However, fluctuations due to changes in scattering of the radar energy from the surface of the bird as if flies through a radar beam alter the periodicity of the amplitude modulation and complicate its use as a reliable signature of identity (Torvik et al., 2014). A fundamental piece of information provided by radar is distance to target, which is not available with a single video camera. Radar also provides much greater range than current thermal infrared cameras, but is generally more expensive and requires more power to operate making it difficult to employ at offshore locations that lack infrastructure. Further, target detection with marine radars is challenging due to noise and clutter (Jarrah et al., 2012), and radar echoes off wind turbines could be problematic if operated within the confines of an offshore wind farm.

We are researching the use of thermal infrared video to capture the flight tracks of birds and bats. The use of thermal imaging to detect wildlife may be limited by excessive humidity, distance from the camera, field of view, physical obstruction, and even plumage and pelage characteristics of the animals being studied (Cilulko et al, 2013). However, although optical cameras have a higher resolution than thermal cameras, thermal video can record observations both day and night and in weather conditions that make optical cameras less effective and is relatively inexpensive to collect and easier to interpret compared to radar. We assume that the shape of the flight path and statistics on changes in direction, which are less affected by low resolution than the image of the animal, will contribute to the classification of a target. Normandeau (2014) recorded birds and bats with thermographic cameras. Flight trajectories and animal shape were used to identify objects as foraging bats, bat/bird, or unknown. Erratic flight trajectories were identified as foraging bats, but it is unclear if or how timing and degree of changes in flight direction were quantified or otherwise evaluated. Although the flight path alone does not function as a species or even group level signature, the automated extraction of data from sensor-derived information could provide a method of significant cost savings for related research. The use of flight path as an indicator of identity, even at a coarse scale, will be a step towards eliciting finer scale classification.

Our objectives are to show that it is possible to automate the identification of behavioral flight paths and to classify the associated birds and bats. We present preliminary models for flight track and target classification based on a limited annotated library of tracks and associated characteristics extracted from a single-camera thermal video. Ultimately, our approach will be useful in estimating or validating the magnitude of risk from blade/tower collisions and evaluating avoidance behavior associated with taxa of concern in areas under consideration for wind energy development.