Survey site and study species

All acoustic survey data were collected at Los Amigos Biological Station and the adjacent conservation concession of the same name, which contains approximately 145,000 hectares of lowland primary rainforest in the province of Madre de Dios in southeastern Peru. Recording was conducted on the station’s existing transect grid, covering an area of approximately 4000 hectares, which was previously used for camera trap studies [56,57]. Habitat heterogeneity in this region is primarily the result of successional gradients associated with increasing distance from the channels of large rivers [58], and is coincident with high avian alpha diversity (~600 species [36]). The two high-level forest types present in the southwestern Amazon are floodplain forest (flooded for at least part of the year) and terra firme (never flooded). Differences in time since disturbance, soil moisture, and edge effects are responsible for creating distinct understory vegetation characteristics in these two habitats as well as the ecotone between them [58,59]. Members of several avian clades regularly segregate spatially by forest type, suggesting that high habitat and species diversity metrics are related [60–62]. eBird survey data were collected from a much broader area within Madre de Dios, some of which is covered by intact forest but elsewhere contains open lands and areas of anthropogenic land use. Madre de Dios currently experiences significant threats from deforestation and fire, and the area bordering the Ruta Interoceánica, a newly-constructed east-west road connecting Madre de Dios and the Peruvian Andean provinces with the Brazilian state of Acre, is a major deforestation hotspot within Peru [63]. Madre de Dios has also become a hotspot for illegal artisanal gold mining in recent years; direct removal of vegetation near rivers and oxbow lakes and large inputs of mercury into the environment have yielded lasting and widespread land use change in this region, even deep into otherwise intact forest areas [15,64,65].

We focused our analyses on a set of five suboscine songbirds in Furnariida, an extremely speciose clade in Amazonian forests: Formicarius analis, F. colma, Myrmothera campanisona, Akletos goeldii, and Oneillornis salvini [36,66] (Table 1), all of which are well-represented in acoustic monitoring data available for the station. These species are all terrestrial or semi-terrestrial, cryptically-colored, and have strong affinities for intact forest with heavy understory vegetation [67], making them important indicators of forest quality [68,69]. At the same time, these species are quite distinct from one another in terms of their utilization of the region’s key ecological gradients, and these differences are well-documented in the literature (Table 1). We used these known differences to our advantage in this analysis by assessing the extent to which they were recapitulated in our candidate models.

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. List of species used in analyses.

https://doi.org/10.1371/journal.pone.0327944.t001

Acoustic data collection

Acoustic data used for model training were collected during three recording seasons. The first two seasons represent existing data from a pilot study and subsequent full monitoring project for tinamous (Tinamidae), respectively [76]. For these projects, 10 Cornell Lab of Ornithology Swift autonomous recording units (ARUs) were cycled across 34 survey points covering the terra firme-floodplain gradient (Fig 1) from late July to early October 2019 (dry season in southeastern Peru) and then again at the same sites from early January to late February 2020 (wet season) [56]. Recording durations were 2–23 days per site during the pilot study and 21 days per site during the wet season. The third season ran from mid-May to mid-September 2024, focusing on 19 new recording sites that were located along the same trail system. These new sites were chosen to provide better spatial resolution within the ecotone between terra firme and floodplain (Fig 1). From May to mid-August, 4 Frontier Labs BAR-LT recorders were deployed for three 30-day long deployments at 4 sites. Starting in mid-August and continuing through the end of the season, 6 Cornell Lab of Ornithology SwiftOne ARUs and 9 Frontier Labs BAR-LT recorders were deployed for an additional single 21-day long deployment at the remaining 15 sites. ARUs were placed no closer than 150 m apart, as this is believed to be the cone of acoustic detectability in dense tropical forest [21]. Recorders were placed ~1 m above the ground with the microphone facing towards the ground, which is optimal for recording understory (most activity ≤ 2 m above ground) and terrestrial birds. Additional details of recording site selection, placement, and recording parameters for the three seasons are available in citation [14] as well as in S1 File.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. Spatial distribution of datasets.

eBird checklists (circles) and recording sites (triangles) within Madre de Dios (A,D), at Los Amigos Biological Station (B,D), and near Puerto Maldonado and nearby agricultural areas (C,F) with respect to the height above nearest drainage (HAND) gradient (A-C) and land use types (D-F). Recording sites were selected to span the HAND gradient (B), while eBird sites occur in a diversity of habitat types. The western yellow box in (A,C) denotes extent of (B,D), while the eastern yellow box denotes extent of (C,F). Study region is outlined in inset map of Peru in (A).

https://doi.org/10.1371/journal.pone.0327944.g001

Processing of acoustic data

Survey audio was captured as a series of WAV files (see S1 File for details on file length) spanning each day’s recording period. Processing was performed using BirdNET-Analyzer version 2.4, which is pre-trained on a list of ~6000 bird species that includes our five target species and much of the avifauna of Amazonian Peru [77]. The classifier examines 3-second moving windows spanning the source audio and returns possible detections with associated detection probabilities (see S1 Table for full classification parameters). A characteristic of all probability-based moving window classifiers is the need to set a minimum threshold for accepted detections. To accomplish this, we performed a semi-supervised calculation of thresholds for each species based on literature best practices for BirdNET confidence score interpretation [78]. We took a sample of the positive detections for each species spanning the full confidence score range of the classifier output (0.3–1); in order to more evenly cover the spatial gradient at Los Amigos, we stratified this sample such that up to 30 detections were chosen from each 10 m interval of elevation across acoustic sampling sites (200–293 m). Species with low average BirdNET classifier performance may have many more low confidence score detections than high confidence score detections, thus reducing sample size in the range of scores where false positives transition to true positives [78]. To account for this, we added a second independent sample of up to 30 detections at a range of higher scores (0.8–1) from each 10 m elevational band. We labeled these detections as true or false using the annotation function of the Raven Pro software suite [79], and fit three candidate models to these labels and their associated logit-scaled confidence scores: i) a fixed-intercept null model, ii) a generalized linear model with logit score as the predictor and label as the response, and iii) a logistic generalized additive model (family “binomial” with REML optimization) with logit score (linear term) as a predictor and label as the response, but with the addition of site elevation (smooth term, cubic regression spline) as an environmental covariate to account for the possibility that classifier performance might vary across the habitat gradient due to the presence of similar-sounding organisms across certain sections of the gradient. Candidate models were evaluated using the Akaike Information Criterion (AICc); for all five species analyzed, the best performing model was the logistic GAM (S2 Table). The threshold for accepting detections was calculated for each species as the lowest (e.g., most permissive) threshold that kept the false positive rate (fpr) as evaluated on each species’ annotated sample set below 5% [78]. Acoustic detections for each species were assigned probability scores using the best fit model and filtered using the chosen threshold. We performed this thresholding process separately for the 2019–2020 seasons and the 2024 season to better account for differences in recording quality and microphone SNR between years, and possible increases in ambient noise at some sites in 2024 due to increased mining activity along the nearby river. In addition, we temporally filtered the 2024 detections so that they matched the diel recording schedule of the 2020 season to ensure greater consistency in survey effort (S1 File). After all other sampling evenness filtering was performed, we collapsed sequences of vocal events into at most one detection per species per hour of recording time [62,80]; we refer to these as “independent detections.”

To facilitate joining acoustic and eBird data for use in model creation, we formatted the acoustic detection CSV file so that it contained the same columns as the eBird dataset download. Using the ‘auk’ package in R [81], we compared the detection set to a database that contained a list of all one-hour recording periods across the field seasons with associated site-date information. Each hour-long recording period received a 1 for each of the target species that was present and 0 for each of the target species that was absent. We hereafter refer to this type of presence-absence dataset taking into account all sampling events as “zero-filled,” as this is the term used in the literature on eBird SDM [82].

eBird data and landcover covariates

For model training, occurrence data for the five target species through September 2024 were acquired from the eBird data portal [35]. Data were pre-processed using the ‘auk’ package in R [81] to select checklists containing at least one of our five target species that were submitted in Madre de Dios between 2010–2023 using semi-structured protocols (Stationary or Traveling). To further increase sampling effort evenness within the dataset, we filtered this dataset to contain only complete checklists (those where the observer specifies that they reported all species they were able to identify) with survey duration less than five hours, effort distance less than 1 km (for traveling checklists), and number of observers less than 5. While these filtering constraints are broadly consistent with what is typically used in the literature on eBird SDM [82], our shorter distance cutoff (1 km as opposed to 5 km) accounts for the fact that ecological gradients in the Tropics tend to be shorter than in the temperate zone [83]. Following the literature on eBird SDM [82], we recoded all checklists with “X’s” for our target species as NAs, which removed them from the dataset. We repeated the previously-described zero-filling process by joining the eBird checklists to the complete database of eBird checklists from the same geographic area, time period, and effort parameters such that each checklist received a 1 if the target species was present and 0 if the target species was absent. Although the set of absences in this dataset could include events where the species was undetected despite actually being present, eBird data are generally treated this way in the SDM literature as they are considered to be a sufficiently good approximation of actual presence-absence processes [84]. In addition, we consider our target species to be very easy to detect by eBirders: as well as having loud, simple vocalizations, they are also among the most common bird species in Madre de Dios (reported more often than 98.0% of all other species in Madre de Dios for F. analis, 87.5% for A. goeldii, 75.4% for O. salvini, 74.2% for F. colma, and 64.0% for M. campanisona). In any case, the measures of occurrence used in this analysis either incorporate detectability indirectly through covariates (encounter rate) or directly through repeated site samples (occupancy), so we felt this issue is of limited enough scope to consider our data to be presence-absence.

Prior analyses using eBird data for species distribution modeling [82] have typically used the Moderate Resolution Imaging Spectroradiometer (MODIS) land-use/land-cover dataset [85] to predict species habitat affinities, with elevation as an auxiliary continuous predictor variable. The MODIS land cover dataset lumps most Amazonian forest types into an “evergreen broadleaf forest” class, and has a resolution (500 m) that is too coarse to resolve the short ecological gradients present in the region [56,86,87]. As a result, we felt it necessary to develop a custom predictor dataset that is resolved to a finer spatial scale and contains more nuanced land cover classes that are relevant for analyses in the Amazon. In order to better capture the natural gradient that exists between terra firme and floodplain forest types in lowland Madre de Dios, we chose to base our analysis mainly on a height above nearest drainage (HAND) surface [59] calculated from the SRTM+ 30 m Digital Elevation Model (DEM) [88] and a linear feature map of the main river channels in the region [89]. We chose to use HAND instead of elevation since, as the primary habitat gradient in this region is maintained by fluvial processes, HAND was likely a more direct proxy for flooding intensity than elevation and accounts for the increase in mean riverbed elevation as one moves west across our study region towards the Andean foothills [56]. We based our categorical land cover predictor set on a Sentinel-2 10 m Land Use/Land Cover Time Series dataset that contains 11 land cover classes (Fig 1) [90]. This land cover layer includes several non-forest land categories (such as bare ground, crops, and built areas); while none have non-forest values at the acoustic points (preliminary data analysis indicates that all of the acoustic points are ≥ 1 km from any non-forest land cover classes), we believed they would be crucial for measuring absences for eBird sites in non-forest areas. Ground-truthing of the Madre de Dios corridor also indicates that dredge tills and temporary habitations on river beaches subject to alluvial gold mining [15] are typically (correctly) classified as “built,” and we felt including these categories might also help our models capture the effects of this form of disturbance on species presence at eBird sites along rivers. We enlarged the size of class four of the Sentinel-2 dataset (“Flooded Vegetation”) by superimposing two land cover classes (“occasionally open water” and “always inundated”) from a 30 m resolution synthetic aperture radar dataset of forest inundation [91] to create a more all-encompassing “flooded vegetation” land cover class. We also included a canopy height predictor based on data from the Global Ecosystem Dynamics Investigation (GEDI) LIDAR project and the Sentinel-2 dataset [92]. All raster values were resampled (nearest neighbor for categorical data, bicubic for continuous data) to 30 m for consistency.

Species distribution modeling is sensitive to spatial scale, and inferences become less robust with increasing distance in spatial and environmental space from the region encompassed by model training data [93]. To ensure that predictions would not be inferred at sites far away from the original sampling locations, we constructed our model prediction surface by creating buffers around each eBird and acoustic monitoring point that we merged into a single polygon and clipped to the borders of Madre de Dios department. We used 35 km as the buffer width for this step as this was the smallest distance which ensured that only a single polygon was left after merging. In addition, we clipped this region with a polygon containing all areas of Madre de Dios with elevations ≤ 350m. All occurrence data used in model training were filtered to exclude points outside this prediction surface, leaving us with data collected at sites with similar elevations and climates to the area where acoustic monitoring was conducted. All GIS layer manipulation was performed using QGIS version 3.26.0 [94].

Modeling framework

Our modeling approach follows that used in earlier studies of eBird data pooling [42,43,82] unless otherwise noted. Although eBird data can be used to model abundance [82], we chose not to construct abundance models in this analysis because our acoustic data are not well suited to capturing this metric. We created three general “classes” of models for this analysis: acoustic only models, eBird-only models, and models containing data from both sources, the latter of which were created by row-joining the two datasets and adding a column labeling each row as an acoustic or eBird observation. All model processing was performed in R version 4.2.0 [95].

Encounter rate

Encounter rate (sometimes also called “encounter probability”) is a commonly used metric in the eBird SDM literature [42,84]. Although it is not capable of isolating the factors that determine site occupancy (z), it is nonetheless a useful approximation of this quantity as it is approximately equal to the product of the site occurrence probability (ψ) and detectability (p) processes from occupancy modeling and uses observation date, time, duration, distance covered, and number of observers as covariates in order to account for between-site variability in detection [82,96]. Model outputs are read as the probability that an average observer will encounter the species given one hour of survey effort during the time of day at which the target species is most frequently recorded [82]. To produce encounter rate models for our species and evaluate the benefit of including structured survey data from the acoustic dataset, we used a bootstrapping approach to fit models on random samples of audio, eBird, and pooled data (n = 100 per model type per species). In each bootstrap iteration, the eBird dataset for our target species was spatiotemporally subsampled by overlaying the sample area with a 600 m x 600 m grid and sampling one observation per week within each cell. This particular grid size was chosen as it is the smallest number that ensures that eBird sample points are no more clustered than the acoustic survey points (mean nearest neighbor distance = 599.34 m). Positive and negative detections were sampled separately to reduce class bias [18]. Spatiotemporal subsampling was not performed on the acoustic survey data because our strategy of placing recorders at set intervals along transects inherently reduced spatial bias. For the set of iterations using pooled data, we combined the two datasets by treating each acoustic sampling event as a pseudo-checklist submitted under the Stationary protocol with number of observers = 1 and duration = 1 hr (accounting for the aforementioned lumping detections into 1 hour bins to reduce pseudoreplication) [42]. Next, we determined the target species’ detection frequency within the sample and, if it was < 25%, used synthetic minority class oversampling (SMOTE) [42,97] to increase the number of presences to ~25% of the total number of observations. For pooled models, the detection frequency calculation and minority class oversampling procedure was performed to eBird and acoustic data separately. Finally, in order to reduce the effect of data pooling and/or oversampling procedures on sample size differences between the three model classes, we performed the bootstrap procedure on the eBird dataset first and subsampled the data used in the acoustic and pooled bootstrap models so that all three models within the same bootstrap iteration had equal sample sizes.

Habitat predictor data were extracted for all spatial points based on raster values within circular buffers around each point. For continuous predictors, mean and standard deviation summary values were calculated within each buffer; for categorical predictors, we calculated percent coverage of each land cover class and edge density (total edge length of patches of the target class per unit area [82]). We first calculated summary statistics for a 300 m buffer radius, or approximately double the radius of the zone of acoustic detectability (150 m) for ARUs in intact forest. We believe 300 m allows for comprehensive measurement of the habitat that is relevant to our target species as it is the distance limit for normal movements of interior forest birds in the Neotropics [68]. As our eBird data contain traveling checklists of up to 1 km, we calculated summary statistics for a second set of 1 km radius buffers and performed exploratory data analysis to determine the strength of correlation between spatial scales. As many predictor summary values were not strongly correlated between spatial scales (S3 Table), we chose to retain both buffers for the encounter rate analysis. Including both spatial scales is acceptable in this step as the randomForest modeling framework is extremely robust against multicollinearity [98,99]. Effort covariates were also included for each observation, including year, day, and time of day as well as duration in hours, distance covered in km, observer speed in km/h (relevant for traveling checklists), and number of observers.

Encounter rate models were fit to each bootstrap’s training dataset using the ‘ranger’ R package and were subsequently calibrated using a constrained monotonically-increasing GAM implemented in the ‘scam’ R package with uncalibrated encounter rate predictions as predictor and known site occurrence frequencies as response. Using a constrained GAM is recommended as a calibration method in the literature on eBird SDM as it reduces the chance of overfitting the model given that the calibration is being performed using the training data [18,100,101]. We produced separate training and evaluation datasets for each bootstrap iteration by splitting the data into 80%-20% partitions with positive and negative classes split separately to ensure that at least one presence and absence were included in each set. As the goal of this analysis was to evaluate models based on how they performed across the entire survey area, we ensured that all bootstrap models were evaluated on test sets containing both eBird and acoustic data. Models were evaluated using i) Cohen’s Kappa, which measures overall performance in predicting presence and absence based on the test set [102]; ii) sensitivity – the probability that a given presence prediction represents a true presence; iii) specificity – the probability that a given absence prediction represents a true absence; iv) Mean-Squared Error (MSE), which measures the mean of the squared deviations between predicted probability of presence and true presence-absence state; and v) the Matthew’s Correlation Coefficient (MCC), which combines model precision, negative predictive value, sensitivity, and specificity [103]. To examine overall patterns of model predictive accuracy, we produced raster visualizations of each species’ encounter rate mean and standard deviation for the entire study area based on predictions from all 100 bootstrap runs within each model class.

Finally, we used the “importance” functionality within ‘ranger’ to calculate relative predictor importance within each model. We chose to measure predictor importance using the Gini impurity score implementation [104], which describes the mean decrease in model accuracy when the chosen predictor is excluded during the fitting process [82]. While the impurity score calculation is highly computationally efficient, it is important to note that Gini impurity scores can suffer from importance dilution when predictors are highly correlated [105]. Our decision to use a bootstrapping procedure necessitated making this tradeoff; however, other researchers seeking to replicate these methods may be better served by using the permutation importance method, which is slower but somewhat more robust against multicollinearity [101]. Our inclusion of predictors from both 300 m and 1 km spatial scales allowed us to directly test whether one spatial scale was preferred over another in each of the model classes. This was done by fitting a linear mixed-effects model [106] with importance score as response, predictor and the full interaction between distance and dataset as fixed effects, and species as a random effect. We also used these scores to test two predictions about relative importance shifts of habitat predictors between model classes. First, as we view the encounter rate metric to be the product of a “signal” (the occurrence process) which we care about and “noise” (the detection component) that we wish to ignore, we value models in which habitat predictors are assigned high ranks relative to detection covariates. As the acoustic dataset has more consistent levels of survey effort than the eBird dataset, we predicted that habitat predictors would be ranked more highly relative to detection covariates in the pooled models than in the eBird-only models. We tested this prediction using a linear mixed-effects model with importance as response, predictor as a fixed effect, the full interaction of “is habitat predictor” (categorical, yes or no) with dataset as a second fixed effect, and species as a random effect. Second, as we felt acoustic data provide more specific information than eBird data about where species occur within intact forest, we predicted that certain forest characteristic indicators (FCIs) should be ranked higher in importance relative to predictors measuring occurrence differences between forest and non-forest areas in the pooled model class compared to the eBird-only model class. We designated HAND and percent cover of floodplain, transition forest, and terra firme as our FCIs: HAND is the single strongest proxy of forest character in this region as flooding disturbance frequency follows an elevation gradient away from major rivers [58], and the two land cover variables are both very strongly correlated with HAND (r ≥ 0.8). As such, we feel species responses to these predictors should strongly indicate in which part of the intact forest landscape they occur. By contrast, we considered canopy height (not strongly correlated with HAND, r ~ 0.4), percent cover crops, percent cover built area, and percent cover bare ground to be more relevant to differences in species occurrence between forest and non-forest areas. We assessed this relationship using a linear mixed-effects model with importance as response, predictor as a fixed effect, the full interaction of “is FCI” (categorical, yes or no) with dataset as a second fixed effect, and species as a random effect.

Occupancy modeling

Occupancy modeling involves jointly modeling occurrence (ψ) and detection probability (p) as separate processes based on relevant subsets of predictor variables [20]. Similar to the encounter rate models, we filtered this dataset to contain only complete checklists with survey duration <5 hours, effort distance <1 km, and number of observers <5. eBird data can be used for occupancy modeling, but in order to meet the standard set of occupancy model assumptions, it requires special processing to select a subset of the data that represents repeat visits by the same observer to the same site within periods of closure over which species presence is believed to be constant [20,82]. We used the function “filter_repeat_visits” in the auk R package [81] to generate a subset of the eBird dataset containing between 2 and 10 repeat visits from the same observer to the same site within a period of 90 days, a period approximately equal to half the length of the two climatological seasons in the southwestern Amazon. Although we are not aware of any literature evidence documenting significant seasonal shifts in habitat utilization for our target species, this has been documented for other species of birds at Los Amigos between peak dry season and peak wet season [62], and we wished to have one closure period for each of these peak periods in addition to one each for the two transitional periods centered on November and April. To account for close spatial clusters of eBird points (typically seen when personal locations are submitted near but not at an eBird hotspot), we rounded all coordinates to 2 decimal points (~1.1 km at this latitude) in the filter function. Spatiotemporal subsampling was again applied to eBird data using the same process used for the encounter rate models. Acoustic data were split into 90 day periods of closure with repeat visits defined as consecutive 3 day-long periods of recording time [80], but were not spatiotemporally subsampled. Data were formatted into site-state matrices using functions from the R package ‘unmarked’ [107]. We calculated the amount of anthropogenic habitat within each buffer size class as the sum of “crop” and “built area” land cover types as these are the two classes that are unambiguously the result of human influence. This accounts for residual differences in occurrence due to human land use that are not explained by canopy height alone. At this stage, we randomly selected 20% of presence sites and 20% of absence sites to hold out as part of an evaluation set, ensuring that both sets contained proportional samples of sites from natural and anthropogenic sites and, for the pooled models, of acoustic and eBird sampling sites. All predictors were mean-centered and standard deviation-scaled as part of the model formula specification.

Occupancy models were constructed using two single-species occupancy frameworks in the spOccupancy R package, both of which are based on modeling Pólya-Gamma latent variables using an MCMC sampler [108]. Audio-only and eBird-only models were fit using the simple occupancy model function “PGOcc,” while the pooled model was fit using the integrated modeling function “intPGOcc,” which calculates joint occurrence probabilities using data from two or more datasets with potentially different detection processes. Occupancy models are extremely sensitive to collinearity, so it was necessary to implement a comprehensive model selection process to select optimal predictor sets. Based on our exploratory data analysis and an examination of preliminary results from the encounter rate modeling component, we allowed audio-only and pooled models to contain habitat predictors from both 300 m and 1 km spatial scales, while only including 1 km predictors in the eBird-only models. For detection covariates, we allowed the eBird-only models to contain a complement of typical eBird checklist effort covariates (day of year, hours of day, effort hours, protocol type (stationary or traveling), effort distance (km), and number of observers). For the audio-only models, we included the subset of these covariates that are relevant to ARU data (day of year, hours of day, and effort hours). To account for the fact that site-surveys represent multiple days in the ARU data, we defined

hours of day as representing the hour in which the first detection of the given species occurred during a sampling period, and effort hours as the total number of hours during each three day survey period that the recorder was turned on. For each model class, we removed all predictors with non-zero variances as they have no intrinsic meaning for the given localities (e.g., anthropogenic habitats among the interior forest acoustic sites) and selected non-correlated predictors using a hierarchical clustering approach [109]. Within each cluster of highly correlated predictors (|r| < 0.7), we selected the predictor with the highest mean importance score in the encounter rate model for the corresponding species and model class. We then performed a multi-stage model selection procedure that has been used by other researchers fitting occupancy models with large sets of predictors using spOccupancy [110]. In summary, we fit each predictor (as both linear and quadratic terms to account for possible unimodal responses) to the training set as a univariate model, with an intercept-only detection formula if the predictor was an occurrence predictor and vice versa if the predictor was a detection predictor. We considered any predictors with a 95% Bayesian Credible Interval (BCI) that did not overlap zero in its respective univariate model to be unambiguously informative. All predictors that were selected in this step were entered into a global model that was subject to backwards stepwise elimination until no additional improvement to Widely-Applicable Information Criterion (WAIC; an extension of AIC that is better suited to systems with latent variables [111]) was observed upon removing more predictors. Predictors in the best model were considered significant if their 95% BCI did not overlap zero, and any that were present in the best model while not being significant on their own were considered minorly informative. Each species-model class combination was fit 10 times using different samples of training data to ensure that results were broadly generalizable. All modeling was performed using default initial values and vague Gaussian priors for beta coefficients with x̄ = 0 and σ² = 2.72. Models were run for three chains, each with 28000 iterations, a burn-in period of 3000 iterations, and a thinning rate of 0.25, and model convergence was assessed by ensuring R-hat values for all parameters were < 1.1.

To assess model fit and compare different model classes, we calculated WAIC for the best model for each bootstrap. We compared model WAIC to the WAIC value of a model with intercept-only occurrence and detection formulas (hereafter “null” model) fitted on the same training data to gauge whether models encoded meaningful information. We also chose to evaluate model performance on an independent evaluation dataset, as this provided a better assessment of model generalization and allowed us to conduct direct comparisons between the three model classes. The core of this evaluation set was the 20% hold-out set from model training; however, we chose to add the set of eBird sites that were removed during the spatial subsampling component of training to ensure that we would still have sufficient numbers of presence and absence points after subsequent filtering steps. Spatiotemporal subsampling was applied to the eBird evaluation sites using the same process as was used during training; we did not spatially subsample the acoustic sites. We extracted our suite of environmental and effort data for site visits and, to match the format of the training data, we filtered visits so that eBird site detection histories represented a random sample of 2–10 repeat visits by the same observer within a period of 90 days and acoustic data detection histories represented 3 day-long periods of recording time.

When evaluating occupancy model performance on new sites, it is important to be reasonably confident that site absences are due to non-occurrence rather than non-detection, especially in situations when survey effort is highly variable across sites [43]. Therefore, we used each species’ single data source models to predict cumulative detectability across all chosen visits using the formula:

Here, p_i is the ith site, t = 1…n is the nth visit to that site, and p_it is the estimated detection probability for each site-visit. Sites with low predicted levels of detectability (≤ 0.9) were removed from the validation set [43]. Although acoustic sites had substantially higher sampling intensity on average than the eBird sites (mean total effort-hours of 176 per site versus 3.6 for the eBird sites), exploratory data analysis nonetheless showed that cumulative detectability at a small number of acoustic monitoring sites with the shortest total recording durations was ≤ 0.9.

The resulting datasets were heavily imbalanced in favor of the eBird sites in all species, so we bootstrapped model evaluation by calculating a set of model metrics (precision-recall AUC, sensitivity, and specificity) on 100 subsamples of the evaluation set that contained all acoustic sites and an equal number of randomly chosen eBird sites with representative ratios of presence and absence sites and reported mean values of the model metrics across all bootstraps. Model coefficients, confidence intervals, and effect size estimates were extracted for comparison between model classes. We also visualized model predictions by producing raster maps of occupancy probabilities for each species-model combination across the study area. Finally, we produced occupancy raster maps by taking the mean and standard deviation of the rasters from each of the 10 bootstrap runs.

Research ethics and organismal impacts

Work was conducted with permission from the Servicio Nacional Forestal y de Fauna Silvestre (SERFOR, permit # D000117-2023-MIDAGRI-SERFOR-DGGSPFFS-DGSPFS). As acoustic monitoring was conducted at remote sites in the rainforest away from populated areas, we are confident that no audio of human speech or specific human activities was collected. Minor exceptions include distant motor noise of boats traveling on the river and dredgers operating at mining sites >2 km away from the station. All data collection was passive, and we do not have any evidence that autonomous recording devices influenced the behavior or well-being of organisms in any way.