Elephant Censusing via Geophone Arrays: A Visual Approach for Linear Arrays Gabriel J Chandler * , Ozgur Izmirli † , Caitlin O’Connell-Rodwell ‡ , and Jason Wood § * Department of Mathematics, Connecticut College, New London, CT 06320-4196 USA, Email: gjcha2@conncoll.edu † Department of Computer Science, Connecticut College, New London, CT 06320-4196 USA ‡ Department of Otolaryngology, Stanford University, Stanford, CA 94305-5739 USA § The Whale Museum, Friday Harbor, WA 98250-0945 USA Abstract—The problem of censusing elephants via geo-phone recordings of their footsteps along a linear array positioned along a path leading to a watering hole is considered. We propose an intuitive and graphically motivated technique that delineates passings of herds of elephants. Additionally, a measure of dispersion of the localization estimates is proposed that correlates to the size of the passing herd. We illustrate the methods using both real data and a simulation study. I. I NTRODUCTION We consider in particular the problem of censusing elephant populations in forested regions of their habitat [Woods, et al. 2005]. The general problem of censusing includes problems of localization, tracking and segmentation of groups of gaited animals in their own habitat. In this work we focus on a particular case where groups of elephants move along a linear sensor array embedded in a noisy environment . The method- ology presented below addresses the problems of localization, tracking and segmentation, and in addition gives indication to the size of the passing group. Other applications include tracking passage of herds to analyze behavioral patterns and time segmentation to extract sections of noisy data for further analysis such as footstep detection and gait modeling. An accurate estimate of the size of an elephant population requires the ability to estimate the number of elephants passing in a group, as well as classification of passing animals as either elephants or otherwise. In this paper, we concentrate on the methods of segmentation of long term data into sections which consist of information pertaining to either a single animal or a group. With proper segmentation, problems such as counting footsteps and classification of animal type (see [1]) is possible, though beyond the scope of the present work. We premise that any method used for solving such problems should be done using the entire array data. As such, we look to segment the data into sections for which the entirety of the array is being dominated by a single group of elephants. We use seismic detection for minimal human intervention to the natural environment. There has been significant research in the field of localiza- tion with a wide variety of applications ranging from acoustic to radar problems. The nature of the sources and properties of the environment dictate different approaches to the detection of the required information [2]-[4]. In source localization terminology, our problem can be viewed as localization of a group of narrowband, near-field, moving seismic sources with unknown signal propagation speed in a reverberant environ- ment. In addition, the sources are impulsive and not necessarily strictly periodic. In our case, the main difficulty in localization arises from seismic properties such as unknown reverberation patterns and rapid energy dissipation. We refer to a small sampling of work related to our method in this paper. Knapp and Carter proposed a generalized correlation method for two-channel time delay estimation [5]. The method used a maximum likelihood estimator and a prefilter to suppress noise power. Mainwaring et al. used a tiered sensor network architecture for habitat monitoring of seabird nesting behavior [6]. Gannot and Dvorkind proposed a two-level method for acoustic source localization with the first level estimating pairwise time delays and the second level combining this information for a microphone array [7]. Other research on acoustic source localization concentrates on arrangement of microphone arrays [8], [9] and tracking multiple speakers [10]. Section 2 contains a description of the data and preprocess- ing, Section 3 describes the procedure for localization, section 4 discusses the parameterization used in estimating location, and results related to estimating group size are given in section 5. A conclusion of the methods and results is found in section 6. II. PREPROCESSING THE ARRAY DATA The input consists of 13 channels of sampled data obtained from equidistant geophones. Due to the fact that the data was extremely noisy we applied a spectral filter that would maximize the ratio of the energy output for footstep signals to all other spectrally unrelated ones. Each channel is filtered in the spectral domain with a filter that has a bandwidth spanning a typical footstep spectrum. This is implemented by calculating the sliding short-time Fast Fourier Transform of each channel. Each window that is 0.25 seconds long is multiplied with a Hann window prior to calculating the transform and successive windows are overlapped 87.5% in time. The resulting banded amplitude spectrum is then squared and summed over all bins to obtain the filtered signal energy for that window. The large dynamic range of the energy signal led us to consider compression in order to reduce the dynamic range. This was more useful for us to visually observe the