Classification of flowers by bats: comparison with the radar case Alessio Balleri and Karl Woodbridge Dept. of Electronic and Electrical Engineering University College London London, UK {a.balleri, k.woodbridge}@ee.ucl.ac.uk Chris J. Baker College of Engineering and Computer Sciences ANU Canberra, Australia chris.baker@anu.edu.au Marc W. Holderied School of Biological Sciences University of Bristol Bristol, UK marc.holderied@bristol.ac.uk Abstract— In recent years, with the development of high range resolution radars, the desire to be able to identify targets under all weather and clutter conditions has become of great importance. This is an activity carried out with great success by echolocating mammals such as nectar feeding bats that are able to detect and select flowers of bat-pollinated plants even in a dense clutter environment. In this paper data consisting of acoustically generated high range resolution profiles of four bat pollinated flower heads are analysed. Multi perspective classification performance is assessed and compared with the radar case. There are close parallels that suggest lessons can be learnt from nature. I. I NTRODUCTION In recent years, with the development of high range resolu- tion radars, the desire to be able to identify targets under all weather and clutter conditions has become of great importance and there is a lot of interest in this and in developing and optimizing radars in order to improve classification perfor- mance. Although, there have been a number of studies on this topic, accurate and reliable classification of targets remains a key problem in many military applications of radar and sonar [1] [2] [3] [4]. This is an activity carried out with great success by echolocating mammals such as bats that are able to detect, select and attack prey even in dense clutter and hostile countermeasures environments. Although the frequencies and waveform parameters used by radar sensors and by echolo- cating mammals are not the same there remain close parallels that suggest lessons can be learnt from nature [5]. The ratio between the targets size and the variation in the wavelength of the transmitted signal is for example comparable in the two cases. Bats have evolved echolocation as a means of detecting, selecting and attacking prey over a period of more than 50 million years into a highly sophisticated capability on which they depend for their survival [6] [7] [8] [9]. It seems self evident that there is potentially a great deal that can be learnt from understanding how they use this capability and how this might be valuably applied to radar (and sonar) systems. Although, in recent radar, performance is still far from that obtained by echolocating mammals, there has been little research that aims to learn from nature as suggested in this paper. Bats use a wide range of signal designs in echolocation. Fac- tors such as frequency, bandwidth, pulse interval and intensity are all shaped by natural selection according to environmental features in the bat’s surroundings and according to its foraging activity. A class of organism that is particularly interesting in terms of object classification by echolocation is bat pollinated plants and their flowers evolved to attract nectar feeding bats not only because of their scent and visual appearance but also by their acoustic echo signature. Even if recognition of flower is a very challenging task, because flowers are motionless, silent and mostly grow in dense cluttered environment, bats show remarkable performance [10] [11] [12] [13]. In this paper we analyse a set of data containing acoustically generated high range resolution profiles (HRRP) of bat polli- nated flower heads. We assess classification performance of a Knn classifier when the images of these flower heads are tested and compare the results with the radar case. Results show close similarities between echoes from floral targets and those from classical radar targets. Multi perspective classification performance as a function of SNR is very similar to that obtained when radar targets are tested [2]. II. DATA COLLECTION The data that were processed were provided by the School of Biological Sciences at University of Bristol. These con- tained high range resolution profiles of four flower heads be- longing to four different species; Amphitecna latifolia, Markea neurantha, Crescentia cujete and Vriesea gladioliflora. The flowers to be irradiated were impaled by a long, very thin insect pin mounted at the top of a thin holder placed in the centre of a small turntable. Revolving the turntable allowed irradiation of the objects from all directions in one plane. The front view of the object was adjusted to 0 degrees. A custom- built condenser speaker and a microphone fixed at a distance of 20 cm from the target at the same height as the target object were used. The distance between the centre of the microphone and the loudspeaker was 18 mm. The microphone was placed coaxial to the loudspeaker, approximately 45 degrees laterally above the horizontal with respect to the midpoint of the 2009 International WD&D Conference 1 9781-4244-2971-4/09/$25.00©2009IEEE