RI conducted an acoustic survey of five species of frogs (Allobates talamancae, Colostethus flotator, Colostethus pratti, Silverstoneia flotator, Silverstoneia nubicola) in September 1981 at one locality - Cerro Campana, along a 650m audio transect (Heyer et al., 1994; Lips et al., 1999). These species were common, diurnal, and terrestrial, each associated with streams to varying degrees. The transect was located along a single-track dirt road (8.68212°N, 79.92871°W, 864 masl - 8.68441°N, 79.93233°W, 878 masl), traversing forest and stream habitats. R.I. counted heard males at every 30 steps, avoiding double counting. These counts were conducted in the rainy season, three times during daylight hours. On average, each transect count was completed in 30 min. These data have not previously been published.

We deployed AudioMoths v1.1.0 (equipped with 32GB SD cards in IPX7 waterproof cases) at 19 stations (5 stream stations and 14 terrestrial stations) across three high-biodiversity localities historically known for rare and possibly extinct amphibians (Figure 1). These localities are covered by tropical premontane forests in the mountainous region of the central cordillera in the Cocle and Panama provinces, where naive populations of amphibians first experienced Bd outbreaks around 2005 (Woodhams et al., 2008). Cerro Gaital is a National Monument and Altos de Campana is a National Park, while Altos de Maria is a low-density urban development. Sampling periods were May 17-November 23, 2022, and July 18-August 25, 2024, coinciding with Panama’s rainy season when most frog species breed and vocalize. AudioMoths were installed at breast height in trees and programmed to record 1-minute segments every 10 minutes for a 24-hour duty cycle (144 recordings/day) at 48kHz sampling rate and medium gain. Station coordinates were recorded using a Garmin GPS. Approximately 128,700 recordings were collected and uploaded to the “Searching for Species of Conservation Concern” project¹ for analysis and annotation (Aide et al., 2013; LeBien et al., 2020).

At each station and season where audiomoths were deployed, two observers conducted minimum of one diurnal and one nocturnal survey along a 100m transect, searching approximately 2m on either side of the transect line or stream (Heyer et al., 1994; Lips et al., 1999). All amphibians detected visually or acoustically were recorded, and when possible, captured and swabbed for Batrachochytrium dendrobatidis (Bd) analysis using an MW113 swab five times on each foot, 10 times on the belly, 10 times on each flank, and 10 times on each thigh for a total of 60 passes (Hyatt et al., 2007; Kriger et al., 2006). Swabs were placed into 1.5-mL microcentrifuge tubes and stored in a –20 °C freezer until processing (Klocke et. al 2023). Bd prevalence rates and 95% Bayesian credible intervals (Jaynes and Kempthorne, 1976) were calculated using an online calculator https://www.causascientia.org/math_stat/ProportionCI.html. Difficult to identify frog specimens were photographed and genetically identified following buccal swabs (Advantage Bundling SP, MW100) and analyzed a fragment of the 16S ribosomal RNA gene (Crawford et al., 2012). We placed the swab in a 1.5 mL tube with 30–40 mg of zirconia/silica beads (Cat. No. 11079105z), added 50 µl of PrepMan Ultra Sample Preparation Reagent (Life Technologies, product No. 4318930), and agitated the tube for 1 min in a Mini-Beadbeater-96 (BioSpec, Cat. No. 1001). Then, we diluted the extracted DNA to 1:10 with DNase- RNase-free molecular grade water (G-Biosciences, Cat. No. 786-292) and PCR-amplified the 16S marker with GoTaq Green Master Mix (Promega, Cat. No. M7122), using an initial denaturation at 95 °C for 720 s and 35 cycles of denaturation at 95 °C for 30 s, annealing at 45 °C for 30 s, extension at 72 °C for 80 s, and a final extension at 72°C for 420 s. PCR products were cleaned with ExoSAP-IT PCR Product Cleanup Reagent (Life Technologies, Cat. No. 78201.1.ML), prepared for sequencing using BigDye Terminator v3.1 Ready Reaction Mix (Applied Biosystems, Cat. No.43-374-55) and BigDye Xterminator Purification Kit (Applied Biosystems, Cat. No.43-764-86), and Sanger sequenced on an Applied Biosystems 3500XL Genetic Analyzer. To determine the species, we used the Basic Local Alignment Search Tool (BLAST) to compare the 16S nucleotide sequences against known sequences in the GenBank database (https://www.ncbi.nlm.nih.gov/genbank/), in addition to morphological features.

We assembled a library of template calls 10 out of 14 species of conservation concern known to occur at the localities surveyed (Table 1). We did not have template calls for 4 of these species Craugastor punctariolus, Craugastor tabasarae, Hemiphractus elioti and Pristimantis museosus (Table 1). However, one sound recording from a single captive individual Ecnomiohyla rabborum exists and presents a rare opportunity to evaluate the utility of template-matching approaches for “rediscovering” lost species (IUCN, 2020a; Mendelson et al., 2008). To validate our acoustic monitoring approach for the rare species we first tested detection of common dendrobatoid species (poison dart and rocket frogs) against observer survey results. We uploaded available template recordings for both conservation-priority species and common dendrobatoid frogs present at the study localities (Table 1). We uploaded reference call recordings to their own “reference call site” using the required date time naming format and added the species under the explore tab. We then visualized the reference calls in Arbimon visualizer (0–24 kHz frequency) and created time-frequency bounding boxes around known calls. Template optimization involved analyzing small data subsets, adjusting to include multiple repeated notes or exclude harmonics as needed to improve true positive ratios. Species with long repetitive calls including four to six notes in the bounding box seemed to be optimal and harmonic notes were not generally helpful to include in the template. These templates are publicly available to all Arbimon users. With templates established, we conducted pattern matching analyses using an initial similarity threshold of 0.3 for pattern matching to minimize false positives (LeBien et al., 2020). For species yielding few true positives, we lowered the threshold to 0.2; for species with very high numbers of matches like Andinobates minutus, Craugastor evanesco and Gastrotheca cornuta, we increased it to 0.4 to facilitate manual validation. Two experts (R.I. and B.G.) validated each pattern-matched annotation by examining audio clips, sonograms, and associated metadata (locality, timestamp). The total number of validated calls from the pattern-matching exercise was divided by the number of days of audiomoth deployment to generate an ‘calls per day’ metric to use as a relative index of abundance to compare stations. The validated dataset was exported as a.csv file for descriptive analysis using ggplot in R (R Development Core Team, 2021).

Template-based sonogram analysis effectively differentiated between frog species (Figure 2). ARUs and pattern matching successfully detected four conservation-priority species: Silverstoneia nubicola, Oophaga vicentei, Triprion spinosus, and Ecnomiohyla veraguensis (Table 2, Figure 3). Detections of these four species were acoustic only, except for one juvenile Ecnomiohyla captured during observer surveys and genetically identified as E. veraguensis (Figure 3). Due to a hand malformation potentially affecting survival, this individual was retained in captivity and provided the first recorded call of the species upon maturity, providing a template for our acoustic detection approach. We found no evidence of Ecnomiohyla rabborum, or Atelopus zeteki that are considered potentially extinct, or 4 other species of conservation concern for which template calls were sourced (Table 2).

ARUs detected Least Concern dendrobatoid frogs at twice as many stations compared to the observer surveys (Table 2). Pattern-matching efficiency varied substantially by species, ranging from 35% for Andinobates minutus to 97% for Colostethus panamansis (Table 3). Notably, C. pratti, historically the most abundant dendrobatoid surveyed in Cerro Campana (Table 4), was absent from this locality despite over 2,000 validated acoustic detections at the other two study localities.

Species exhibited distinct calling patterns (Figure 4). Dendrobatoid frogs showed exclusively diurnal activity but with species-specific peaks: S. nubicola displayed crepuscular behavior (peaks 6–7 am, 5 pm); S. flotator showed bimodal activity (peaks 10 am, 4 pm); C. panamansis and A. minutus demonstrated strong early morning activity (6–7 am) with continued calling throughout the day. A. talamancae and O. vicentei calling peaked at midday, while C. pratti exhibited trimodal activity (peaks 8 am, 1 pm, 4 pm). These diel calling patterns of dendrobatoid frogs were mirrored by the historic acoustic call surveys, but they lacked the fine scale temporal resolution of the passive acoustic arrays (Table 4, Figure 4). E. veraguensis vocalized primarily at dusk (7 pm) and pre-dawn (5 am), and T. spinosus called mainly between 7–9 pm but these graphs were based on a very small number of calls, 18 and 10 calls respectively.

The amphibian community showed moderate Bd prevalence (21%, 95% Bayesian credible intervals [14.6%, 28.1%], Table 5), with 251 amphibians from 35 species visually detected across 31 transects with mean abundance of 8.1 amphibians visually encountered per 100m transect. The physical searches represent 66.8 hours of search effort amounting to 3.75 frogs per hour (Table 5). Only frog species were Bd positive with infection loads ranging between 2 to 7,377,000 zoospore equivalents (median = 807 zoospore equivalents, Table 5).

Additional anecdotal observations outside formal sampling in Cerro Campana included observations of T. spinosus and S. nubicola and Agalychnis lemur individuals (Figure 3). A physical specimen of O. vicentei was encountered outside of these surveys near Bajo Bonito (04/30/2025, 8.67582411°N, 80.05116264°W, 905 masl), but it validates the calls detected nearby at Cerro Gaital, especially considering that the IUCN portion of this range was documented as extirpated (IUCN, 2020b). We documented a new locality record for Pristimantis gretathunbergae in Altos de María (05/20/2022, 8.63509°N, 80.07840°W, 1042 masl), based on a recently deceased frog found on the forest floor (STRI Collection of Herpetology, specimen CH-10577) with a Bd load of 63.8 x 10⁶ zoospore equivalents. Several species of conservation concern lacking template calls (Craugastor punctariolus, C. tabasarae, C. evanesco, Pristimantis museosus, Hemiphractus elioti) were also not detected anecdotally, or during observer surveys.

The effectiveness of ARU detection varied significantly among species. Pattern-matching proved highly effective for species with distinctive, repetitive calls like Dendrobatoidea, but less reliable for those with simple or quiet vocalizations (Figure 2). For Ecnomiohyla veraguensis and Triprion spinosus, the use of pattern matching had low ratios of validated pattern matches, but at least they have loud, distinctive and repetitive patterns in their call. This may partly be explained by the fact that they are very rare, rather than the lack of a suitable call for employing this approach. The use of different thresholding for different species was a tradeoff for managing post-pattern matching manual validation effort. While this may raise concerns that very different numbers of positive calls may be found based on those thresholds, one must inherently assume that the detectability of each species using this method is very different. As such, inter-species call frequency comparisons should be avoided, but within a species detectability should be the same as long as thresholding was consistent across all sites. The high false-positive rate for Agalychnis lemur, Craugastor evanesco and Gastrotheca cornuta, who have very simple and loud but monosyllabic calls without a consistently repeated call pattern highlights the need for more sophisticated machine learning approaches as more training data becomes available. For example, machine-learning approaches have been successfully deployed for A. lemur in another study in Costa Rica that used template optimization pattern recognition combined with machine learning to reject false positives (Chirino et al., 2025). However, these require large volumes of annotated training data beyond the scope of this study given the large numbers of species and paucity of training data.

While we failed to detect E. rabborum, our successful acoustic documentation of E. veraguensis validates ARU methodology for sampling rare, arboreal Ecnomiohyla species. The genus’s elusive nature, canopy-dwelling habits, and infrequent calling patterns (Kubicki and Salazar, 2015) suggest continued ARU monitoring may be valuable for detecting remnant populations. One barrier to broader implementation of this method is the paucity of reference calls. While captive recordings exist for several threatened Panamanian species (Gratwicke et al., 2016), and reference calls are available through established sources (Ibáñez et al., 1999; www.fonozoo.com), platforms such as www.inaturalist.org and www.xeno-canto.org are emerging as important crowd-sourced repositories for wild frog vocalizations.

The passive acoustic monitoring method has revealed a fine scale diurnal calling frequency pattern for the first time for several species. Similar efforts to replicate this type of analysis in the past required highly trained experts and an extraordinary field effort (RI pers. obs.; e.g., Navas, 1996). This information is invaluable to design appropriate survey strategies and timing of effort especially for very rare species. For example, S. nubicola is identifiable by its call, but its crepuscular nature does not coincide with the hours when diurnal or nocturnal surveys are often conducted and may explain partially why earlier search efforts did not find them. As research on soundscapes considers the impacts of anthropogenic disturbance on reproduction-related biophony is becoming more relevant to environmental mitigation work (Deichmann et al., 2017).

To make historical comparisons of amphibian abundance, we used visual encounters as the response variable to be consistent with past methodology (Crawford et al., 2010; Kilburn et al., 2010). Current amphibian abundances (3.6-11.3 amphibians seen/100m transect, Table 5) suggest partial recovery compared to post-decline levels (0.2-6.6 frogs/100m) reported by Crawford et al. (2010) from El Copé National Park but these numbers are still well below pre-decline abundances in El Copé (23–32 frogs/100m) (Crawford et al., 2010). When looking at search effort (number of individual amphibians per person hour search), visual encounters survey abundances of 1.7 - 2.8 amphibians/hour of search effort (Table 5) are higher than the 1 amphibian/hour of search effort documented for El Copé two years post-epizootic, but are below pre-epizootic abundances documented for Cerro Campana of 4 amphibians/hour of search effort (Kilburn et al., 2010).

The prevalence of Bd in amphibians has decreased compared to historical records obtained during chytridiomycosis outbreaks in 2006–2007 at Altos de María and Cerro Campana (i.e., 78% and 47%, respectively; Woodhams et al., 2008; Kilburn et al., 2010). Nonetheless, the overall persistence of Bd (21% prevalence) with continued mortality indicates ongoing disease pressure despite the apparent recovery of some populations. While counting calls repeatedly at a single station is not necessarily a good indicator or abundance of animals, it offers a high resolution and accurately measurable relative index of species-specific reproductive effort in a geographic location as a baseline for future comparisons. Establishing baseline sampling protocols, archiving calls and making them publicly accessible through platforms such as Arbimon will be an asset for future recovery monitoring.

Our acoustic surveys documented the persistence of three conservation-priority species: O. vicentei and T. spinosus were presumed extirpated from most these localities following chytridiomycosis-related declines, and S. nubicola was last documented from Campana National Park in 2016 (Crawford et al., 2010; IUCN, 2020c, IUCN 2020d, IUCN, 2020b; Woodhams et al., 2008). The detection of over 400 O. vicentei calls confirms this endangered species’ presence in the El Valle region, where it was thought extinct (IUCN, 2020b). Similarly, T. spinosus recordings provide the first evidence of Panamanian populations in a decade (IUCN, 2020d), though limited detections (10 positive sound files) suggest a small population size. The documentation of S. nubicola, severely impacted by chytridiomycosis at the study localities around 2006-2007 (Kilburn et al., 2010; Woodhams et al., 2008), indicates a potentially recovering population.

Some species that disappeared from Cerro Campana (Woodhams et al., 2008) have reappeared (i.e. Pristimantis caryophyllaceus, Pristimantis cruentus, Pristimantis pardalis, Agalychnis lemur, Smilisca sila, Andinobates minutus, Sachatamia albomaculata, Cochranella euknemos, Allobates talamancae, Colostethus panamensis, Silverstoneia nubicola, Rhinella horribilis) and appear to be recovering. However, C. pratti, remains missing even though it has recovered at other localities. This presents an opportunity to experimentally translocate animals from the other recovered populations of this species that we documented in this survey, which is a recommended intervention strategies for amphibian disease recovery that assumes evolved resistance in the source population (Knapp et al., 2024; Mendelson et al., 2019). These findings also present opportunities for other targeted conservation actions recommended by the IUCN global amphibian conservation action plan (Bletz et al., 2024). Specifically, augmentation with captive releases and in-situ breeding site supplementation strategies that have proved successful in Costa Rica (IUCN, 2020d, IUCN, 2020e). However, any conservation actions should be carried out on a small scale, accompanied by a responsible adaptive management framework with monitoring of host communities and Bd prevalence in order to better understand the effects before being applied broadly (Bletz et al., 2024).

Our findings demonstrate the value of combining passive acoustic monitoring with traditional surveys for detecting rare and recovering amphibian populations. While pattern-matching approaches show promise, particularly for species with distinctive calls, our current methods are limited and underscore the need for expanded call libraries and focus on developing advanced machine-learning techniques. The fact that we do not have template calls for pattern-matching efforts of five species of conservation concern highlights the need for publicly available sources of calls to be archived for research purposes. The documented persistence of previously missing species at our localities in Panama highlights the importance of long-term monitoring and suggests that some species may be developing disease resistance, offering hope for amphibian conservation in the region.