Insects and Recording Procedures

Sounds produced in 6 two-year-old potted date palms infested with RPW larvae were compared with sounds in 3 uninfested controls. Infestation was conducted by inserting a 3rd-4th-instar (85.7 ± 31.6 mg) into a 6-mm-diam. hole bored into the trunk. Sounds of larval activity were recorded weekly for 10 weeks for 1-2-min periods, once per week starting one week after infestation, with a microphone and amplifier (Larven Lauscher, NIR-Service, Bad Vilbel, Germany) attached to each tree (Soroker et al. 2004). In addition, laboratory experiments were conducted to examine patterns of insect feeding and movement sounds. Second and third instars (20-30 mg) were introduced into 28-33-mm-diam. stalks of sugarcane (Saccharum officinarum L), cut into lengths of 60-80 mm, on which they fed readily. The larvae were inserted into holes drilled parallel to the fibers. A 1-2-min record was obtained from each stalk 62-69 d after egg-laying. The larvae were weighed immediately after the recordings had been obtained, and ranged in weight from 0.17 to 1.7 g (mean ± SE, 0.76 ± 0.156 g). Recordings also were obtained from uninfested sugarcane stalks.

For comparisons with other tree-boring insects, sounds of pine sawyer (PSW) larvae, Monochamus titillator (F.), were collected for 1-2-min periods on several different days from a pine bolt with the aid of an accelerometer system (Mankin et al. 2000, 2002). The ages of the larvae were unknown, but 4 adults emerged over a 1-month period, beginning about 2 weeks after the recording. Sounds of 2 buprestid (BUP) larvae, Agrilus dozieri Fisher, were collected similarly from an oak tree branch about 2 weeks before the adults emerged, 3 d apart. The experiments with PSW and BUP were conducted in a laboratory with low levels of background noise. In all the experiments, subjective evaluation of larvae activity was conducted by listening to the sounds with a headset while recording (usually by an author). Recordings were digitized at 44.1 kHz on a computer and saved in .wav format.

Signal Processing

Digitized signals were analyzed with Raven 1.2 software (Cornell Lab of Ornithology, Ithaca, NY) and other dedicated signal sampling software developed in Matlab (MathWorks, Inc., Natick, MA), or with a customized software program, DAVIS (Mankin 1994; Mankin et al. 2000). Signals were band-pass filtered between 0.2 and 12 kHz to eliminate low-frequency background noise. Fast Fourier Transforms were calculated on 256-point time-slices of the waveforms with a Hamming window, and spectrograms were calculated from sections with 90 per cent overlap. Mean spectra (profiles) used for characterizing and identifying putative insect sound impulses were calculated based on 512-point time slices centered on the peak of each impulse in a section of recording independently verified to contain insect sounds without background noise (see Mankin 1994; Mankin et al. 2000, 2007).

In playback and oscillographic analyses of these signals, it became evident that a considerable fraction of the larval sound impulses were of low amplitude, barely above the peak background noise levels. Consequently, it was necessary to systematically adjust the signal processing amplitude threshold, T_a (Mankin et al. 2000), to maintain its level just above background. This was accomplished by calculating the root mean square of consecutive samples in 0.186-s (8192-point) time slices to estimate the peak to peak background noise in each interval. The T_a was reset for each time slice by multiplying the root mean square by a user-settable factor (3.25 in these tests), which held T_a just above the local peak background noise.

The DAVIS program identified and timed groups (trains) of impulses with interpulse intervals less than a preset duration, I_i, storing the beginning and end times of these trains in a spreadsheet along with the numbers of impulses in each train. Impulses that failed to match one or more specified insect sound profiles were discarded and not entered in the spreadsheet. The records contained sometimes as few as 30 or as many as 10,000 impulses, of which about 5%-70% were discarded, depending on the background noise levels. The beginning of a train was set as the beginning of the first valid impulse after a period where the interpulse interval was > I_i, and the end was set as the end of the last valid impulse whose interpulse interval was < I_i. We analyzed the signals in this study using the setting, I_i = 500 ms.

Larval Sound-Impulse Characteristics

Examples of sounds produced by moving and feeding R. ferrugineus, A. dozieri, and M. titillator larvae in a palm tree, oak branch, and pine bolt, respectively, are shown in Fig. 1. Signals produced by RPW in sugarcane were similar in temporal pattern to those in palm. The beetle larvae produced variable-amplitude, 3-10-ms impulses with strong energy between 0.4-8 kHz (see spectrograms A_s, B_s, and C_s in Fig. 1). The signals extend to higher frequencies than the 0.5-1.8 kHz range of sounds produced in soil because wood has a lower attenuation coefficient than soil at high frequencies (Mankin et al. 2000). Other differences between the transmission of sound in a fibrous structure and soil include resonances that are influenced by the dimensions, mass, and stiffness of the structure (Cremer et al. 1988; Evans et al. 2005; Hambric 2006). Signals with energy at a resonance frequency will transmit farther in structures than nonresonant signals, and the long-distance transmission of background noises with components at resonant frequencies can interfere with detection of weak, insect-produced signals. One such resonance can be seen as a continuous, faint band near 2.5 kHz in the RPW spectrogram (Fig. 1A_s), and similar resonances can be seen near 3 kHz in the BUP spectrogram (Fig. 1B_s) and near 4.6 kHz in the cerambycid spectrogram (Fig. 1C_s).

The effects of resonance are observable not only in the background noise spectra but also in the spectra of larval sound impulses (Fig. 2A-B). The examples of RPW sounds recorded from separate palms have peaks near 1.1 and 2.6 kHz in Fig. 2A. The sounds recorded from sugarcane have similar peaks. However, an example of palm tree background noise also has peaks very close to these values, at 1(marked at BP1 in Fig. 2A) and 2.4 kHz (marked at BP2); consequently, the energy at these frequencies is not due necessarily to the larvae alone. This result is unlike what is typically found for background noise in soil or air (Mankin 1994; Mankin et al. 2000; Mankin & Benshemesh 2006), where the energy of most background noise was found to decrease rapidly above 0.2 kHz. In previous studies with soil insects, the relatively low energy of background noise at frequencies above 0.5 kHz enabled construction of insect sound profiles and background noise profiles that could be used to identify individual sound impulses as insect sounds or background noises (Mankin et al. 2000, 2001, 2007).

For RPW in palm trees, there are differences in the spectra of larval and background noise between 3.4 and 6 kHz (Hetzroni et al. 2004a, 2004b, see also Fig. 2A) that can be used to distinguish RPW sounds from background noise. Such differences were used to correctly classify ca. 90% of sounds produced in infested palm trees (Hetzroni et al. 2004a, 2004b). However, the examples of BUP spectra and background noise in oak branches present a more complicated result. In Fig. 2B, the BUP signal has peaks near 4.7 and 5.5 kHz, but there also are large peaks near 4.5 and 54 kHz in background sounds recorded from an uninfested oak branch. Other examples have been observed, particularly in urban environments, where background noise contains peaks of high energy above 0.5 kHz; which provides impetus to develop another method in addition to spectral profile analysis to help distinguish insect sounds from background noises.

Larval Sound Temporal Patterns

An additional method of interest for discriminating insect sounds from background noise is the identification of repetitive patterns that may occur when the larvae move inside their tunnels or scrape the wood fibers during feeding or tunneling activity. An example of such a pattern is seen in Fig. 3, adapted from Hetzroni et al. (2004a, 2004b) A group of repeated impulses appear in Fig. 3Aa that are clearly distinguishable from the unpatterned impulses that were recorded from the uninfested palm (Fig. 3Bb). The existence of identifiable, repeated patterns in records of insect sounds has been noted previously (Andrieu & Fleurat-Lessard 1990; Mankin et al. 1997; Zhang et al. 2003).

Unlike the patterned sounds produced during insect communication (e.g., Walker et al. 2003), the RPW larval sound patterns examined in this study, as well as those produced by the BUP and PSW larvae, were consistent over only short periods of time. The most consistent pattern was not the series of high-amplitude impulses seen in Fig. 3A, but rather a mixture of low- and high-amplitude impulses that occurred in well-defined trains with short interpulse intervals, separated by 0.5-s or longer intervals between trains. Experienced listeners usually recognized these trains as distinct sounds, and frequently recognized them as characteristic of insect sounds.

The numbers of impulses per train varied considerably over the records in the palm tree study (1-99 impulses) and even greater in the sugarcane tests (1-239 impulses). Playbacks of records suggested that trains with >6 but <200 impulses were most recognizable to experienced listeners as insect movement and feeding activity. Generally, background noises in records from uninfested wood or sugarcane had fewer than 7 impulses per train. Consequently, we separated out a particular class of impulse trains, those with >6 but <200 impulses separated by intervals <500 ms, as potential predictors of insect infestation, denoting them as bursts.

The distributions of impulse trains and bursts in six recordings from different insects are shown in Table 1, and examples of several impulse trains denoted as bursts are shown in Fig. 4. Most of the computer-identified bursts in the 4 examples, denoted by the numbered, shaded areas, correspond to groups of low-and high-amplitude impulses separated by periods with only sparsely occurring impulses. In Fig. 4A1, however, almost all of the impulses are low-amplitude, which suggests that the use of burst analysis as a method to augment profile analysis will be most successful in tests where sensors can detect very weak signals near the background noise threshold.

Use of Impulse-Train and Burst Analyses to Assess Infestation Likelihood

An important goal of insect acoustic detection is to discriminate whether or not individually tested samples contain infestations. A successful approach has been to construct indices of infestation based on quantitative and/or qualitative characteristics of detected sounds, and assess the likelihood of infestation using discretized indicators, e. g., low, medium, and high likelihoods of infestation based on lower and upper thresholds of the indices (Mankin et al. 2007). In 1 example, recording sites in soil were classified at a low likelihood of white grub, Phyllophaga crinita (Burmeister), infestation if 2 or fewer sounds (impulse-trains) per min were detected that matched white grub spectral profiles and at high likelihood if greater than 20 impulse-trains per min were detected (Mankin et al. 2007). Here we considered a similar use of impulse-train bursts to assess RPW infestations in the laboratory tests with sugarcane.

Initial analyses of records from infested and uninfested sugarcane indicated that the RPW-in-palm and RPW-in-sugarcane spectral profiles in Fig. 2 readily classified background signals as noise with DAVIS and correctly identified signals produced in infested sugarcane as being produced by RPW. To consider the distribution of impulse-train bursts in infested and uninfested sugarcane, we analyzed one 50-s segment recorded from each of 21 different infested and 4 uninfested stalks, discarding sounds that did not closely match the profiles (Mankin et al. 2007, 2000). The results are summarized in Table 2, where the complete data set is separated into 4 categories, 1 for recordings in uninfested sugarcane, and 3 for records in infested sugarcane where the numbers of bursts ranged from 7 to 17, 4 to 6, and 0 to 3, respectively, and the mean larval weights, numbers of trains, and numbers of impulses per train are listed for each burst category. The correlation the numbers of bursts and the larval weight was not significant (Spearman ρ = 0.41, P = 0.08 for null hypothesis of ρ = 0) (Proc CORR, SAS Institute 2004a). Likewise, the correlation between numbers of impulse trains and larval weight correlation was not significant (Spearman ρ = 0.06, P = 0.79 for null hypothesis of ρ = 0) (Proc CORR, SAS Institute 2004a).

The observed distribution of bursts suggested that the likelihoods of infestation (Mankin et al. 2007) in these tests could be set at: Low, n_b < 1; Medium, 1 ≤ n_b < 3; High, n_b ≥ 3, where n_b is No. bursts per 50-s record. Based on these criteria, 2 of the 4 records from uninfested sugarcane had Low likelihood of infestation, and 2 had Medium likelihood of infestation, while 2, 1, and 18 of 21 records from infested sugarcane had Low, Medium, and High likelihoods of infestation, respectively. These 2 distributions were significantly different under the Wilcoxon Two-sample Exact Test, with sum of scores = 17.0 and P = 0.0025 (Proc NPAR1WAY, SAS Institute 2004b). Although additional testing will be necessary to establish the practical efficacy of incorporating burst analysis into acoustic surveys of RPW in palm trees and offshoots, these results indicate that it is possible, using already existing signal processing software, to extract previously unmeasured information from acoustic recordings that can assist the process of discriminating sounds produced by hidden insect infestations from background noise. The burst analysis method appears to be robust in that it can be applied to signals from multiple species of insects recorded from multiple species of trees. Experienced listeners will always have superior capabilities to identify spectral and temporal patterns that assist in discrimination of insect sounds from background noise, but better understanding of the behaviors that produce the detected sounds and better understanding of the specific methods used by listeners in identifying signals likely will lead to continued improvements in automated acoustic detection methods.