1. Data collection
Data collection was conducted under the auspices of the Amazon Conservation Association at the Los Amigos Conservation Concession (LACC), in the lowland rainforest of Madre de Dios, Peru. This site, which protects ~145,000 ha of forest along the Río Los Amigos basin, is one of the most biodiverse lowland rainforest sites in the Amazon basin with close to 600 bird species, eleven of which are tinamous in the genera Tinamus and Crypturellus (eBird, 2017; Table 1). The station’s biological diversity is due in part to its diversity of terrestrial microhabitats, which include terra firme and floodplain primary forest, secondary and edge forest, Guadua bamboo stands, and Mauritia flexuosa palm swamps (Larsen et al., 2006). As studies at this site (Mere Roncal et al., 2019) and elsewhere in the Neotropics have demonstrated that tinamou species differ in their specific habitat utilization characteristics (Guerta & Cintra, 2014), LACC is an exemplary site for detecting tinamous across a variety of habitat gradients.
Acoustic monitoring was conducted using ten SWIFT ARUs (Kahl et al., 2019), provided by the Cornell Lab of Ornithology, from mid-July to early October of 2019. This period overlaps with the latter half of the dry season at LACC. The SWIFT units were deployed on rotating 14 day deployment periods at terra firme and floodplain forest sites (Figure 1, S1), 10 sites at a time, over three deployments from mid-July to late August. A fourth deployment, duration 27 days, was conducted as a follow-up at five of the 30 sites from late September to early October. As the chosen sites are part of the station’s existing camera trap system (approximately a 1 km2 grid located along the edge of open trails), we were able to merge our detection set with previously-collected site-level habitat data as well as to compare our tinamou detection rates to those calculated using camera trap detections. Recorders were tied to trees at a height of approximately 1.5 m from the ground with the microphone facing downwards. Each unit was programmed to record for five hours a day, from 5:00 to 7:30 in the morning and 16:00 to 18:30 in the afternoon to early evening, in order to cover periods of high vocal activity for tinamous (Dias et al., 2016). The SWIFT unit firmware allows for control of microphone gain and sampling frequency; we set these values to -33 dB (the default) and 16 kHz, respectively. Setting the sampling frequency to 16 kHz is a tradeoff that limits the acoustic frequency bandwidth to 0-8 kHz (Landau, 1967) in exchange for smaller file sizes and lower power demands than the default value of 32 kHz. The SWIFT firmware writes data as 30 min-long WAV files (~58 MB). Each unit was intended to collect data for the shorter of (a) the entire 14 or 27 day recording period or (b) until battery power was exhausted. In practice, battery life was always the limiting factor, with a mean time-to-shutdown of 7.81 days (5.12 days for deployments 1-3 and 21.8 days for deployment 4). Due to supply limitations, we were forced to use a different brand of battery for deployments 1-3 than for deployment 4, which we suspect is at least partially responsible for the longer per-recorder run times in the latter deployment. At the end of each deployment period, all units were removed from the field, loaded with fresh recording media and batteries, and deployed to their next assigned site on the following day. All audio data was backed up to rugged solid state storage media for transport out of the field.
Our chosen classification procedure is a type of supervised machine learning, which requires a significant amount of training audio to produce a working model (Kotsiantis et al., 2007). We used a set of ~3100 audio files of 2s duration (the typical phrase length in tinamou calls) to train an initial classifier. These files were coded as one of twelve classes: one class for each tinamou species, and a “junk” class containing audio of other bird species, non-bird organismal audio, and assorted environmental audio (Table 1). The training dataset was derived from audio downloaded from the Macaulay Library of Natural Sounds (https://macaulaylibrary.org) and Xeno-Canto (http://www.xeno-canto.org) databases (S2) as well as from exemplar cuts in the audio we collected in the field.