When selecting focal species, the annual cycle needs to be considered. The large majority of bird species follow a breeding season cycle (Sharp, 1996; Carey, 2009), and vocal activity can vary throughout the year driven by seasonal differences in calling rates, such as during the breeding season (Catchpole and Slater, 2003). Some species have migratory periods (latitudinal or altitudinal), which result in all or most individuals abandoning a site during part of their annual life cycle (Barcante et al., 2017). This variation needs to be considered in the context of the timing of farm agricultural management practices of interest (i.e., pruning or harvesting) in order to associate changes in calling rates directly with a specific farm management practice (for more details on seasonal transitions, please see Supplementary Materials ).

Duty cycle, the proportion of time that birds vocalize during the day compared to the total time for other activities (Staicer et al., 1996), should also be considered for selecting focal species. Most bird species have two peaks, i.e., an early dawn chorus and a chorus at dusk around sunset, and not all bird species exhibit both types of choruses (Staicer et al., 1996; Bruni et al., 2014; Farina and Ceraulo, 2017). Activity peaks during the dawn and dusk chorus have also been shown to vary depending on the weather conditions, daylight, or temperature (Hutchinson, 2002; Bruni et al., 2014; Marín-Gómez and MacGregor-Fors, 2021). For example, if the morning is cold, cloudy, and with high levels of precipitation, vocal activity has been shown to be delayed for up to an hour, and the timing of ARU recordings should account for that variation. Changes in daylight will also change the timing of vocal activity, even for tropical regions (Staicer et al., 1996; Hutchinson, 2002; Bruni et al., 2014; Marín-Gómez and MacGregor-Fors, 2021).

The differences in species presence and abundance and its fluctuation during the day in agricultural environments will likely influence the detection of certain species. Detectability and vocal activity will be affected also by weather, resource availability, and disturbances (e.g., coffee harvesting) (Thomas et al., 2020; Morelli et al., 2022). Information on natural history and behavior can be used as well as the informed criteria of an expert, but in general, species that tend to be quiet or only vocalize in very specific situations may not be useful, as well as species that can be detected at distances longer than the area of study where the agricultural management practices are being evaluated.

The species selected to evaluate the effects of a given management practice should represent the main foraging guilds (e.g., frugivores, carnivores, nectarivores, or insectivores) that are likely to be influenced by the specific management practice of interest. For example, foliage gleaners are insectivores and are more likely to be affected by the application of foliage agrochemicals such as fertilizers or insecticides, relative to fertilizers that are applied on the ground (Alharbi et al., 2016). The foraging strata (e.g., canopy, lower understory, leaf litter, tree trunks, or branches) should also be taken into consideration, given that species that forage on leaf litter or coffee branches are likely to be more affected by heavy pruning for example (Dietsch et al., 2007).

The main barrier that exists currently is the high level of effort that needs to be invested to carry out all of the annotations needed to detect all the species in a community, especially considering that it has been shown that manual detection has higher accuracy and is more time-efficient, in specific cases like for species that have several song types per individual, and song types varied between individuals and geographically (Venier et al., 2017; Joshi et al., 2017; Cole et al., 2022). Therefore, selecting a small number of focal species will be more informative and more likely to yield robust inferences.

The factors for defining an optimal ARU recording protocol begin by answering the question of when to record; as previously discussed, different bird species in the community are likely to have different responses to the same management practice depending on their ecology and natural history (Vickery et al., 2001; Kirk et al., 2011; Alharbi et al., 2016). If the goal of the study is to understand the responses of a couple of species to agriculture practices, the programming of ARUs should be guided by previous knowledge on the vocal activity of those species. If the focal bird species tends to vocalize mostly during the dawn chorus, limiting the recording to only the dawn chorus will save 50% of storage capacity and battery life by reducing the total recording time from 5 h/day to 3 h/day. An advantage of focusing primarily on the dawn chorus for recording is that weather conditions tend to be more stable and predictable relative to the dusk chorus and daylight hours at sunset tend to be influenced by the degree of cloudiness, ultimately reducing the total sampling period for the dusk chorus (Staicer et al., 1996). In addition, more stable temperatures and lower average wind speeds at dawn have been found to increase the sound transmission of bird vocalizations (Catchpole and Slater, 2003) (For more details on when to record and the suggested duration of the recordings, please refer to the Supplementary Material ).

The number of ARUs per unit of inference is also a matter of efficiency, and one ARU deployed per plot (1,000 m² to 2,500 m²) should be sufficient to monitor the effects of agricultural practices on bird vocal activity If agricultural management practices and conditions differ between plots (e.g., large farms), multiple plots can be monitored. Care should be taken when using more than one ARU per plot, since placing ARUs too close to one another might cause the detection of the same vocalization more than once, especially species that vocalize loudly and can be heard for long distances or species with very large territories (>1 ha). In this case, an intensity threshold can be defined to select which sounds are taken into account when analyzing the data in a spectrogram. However, if the goal is to compare two treatments (e.g., two different pesticide active ingredients applied in similar plots of the same farm), or two vegetation types in the same plot (e.g., the plot matrix and the living fence), then paired ARUs could be used to detect differences in bird activity between those treatments or vegetation types.

Finally, factors to determine how to best analyze the recordings include acoustic data storage. Acoustic monitoring with ARUs requires a significant investment in data management that can often be challenging. The size of a small-scale deployment can easily result in terabytes of recorded audio files. As a result, compression and transfer operations can often last multiple days and require expertise in big data management (e.g., parallel processing). The sound recordings should be first backed up to a single local hard drive as WAV files directly from the SD cards of the ARUs. Second, compressing these files to FLAC 1.4.3 is recommended, a format that maintains the audio properties without data loss and allows a file size compression of up to 70%. This type of compression rate makes it an ideal format for bioacoustics because it is commonly supported by bioacoustics data processing tools (MacPhail et al., 2024). Once the data are backed up in two hard drives using the two different formats (WAV and FLAC), the SD cards can be cleared and used again.

Analyzing large datasets of acoustic recordings from ARUs can be challenging, especially when the interest is being able to detect the entire avian community (Brandes, 2008; Digby et al., 2013; Darras et al., 2019). Overall, community-level analyses can involve a combination of computational methods, signal processing techniques, machine learning algorithms, and an efficient computing system to optimize the duration of processing and analysis (Brandes, 2008; Darras et al., 2019). In addition, analyzing ARU data often involves very specialized personnel, who identify and annotate all or a set of focal bird species present in the recordings, depending on the type of analysis that will be used: a trained detection algorithm or raw annotated data (Digby et al., 2013; Sugai et al., 2019). Both of these approaches can be time-consuming and are two of the main barriers for addressing research questions using ARU data.

The general type of data analysis is an approach most commonly used when researchers are interested in detecting all species in a bird community in the study area. The first step is to preprocess raw audio recordings to enhance the quality of the data and facilitate analysis. Second, audio recordings are segmented into smaller units to facilitate data management and analysis (e.g., Brandes, 2008; Priyadarshani et al., 2018). The last step is classification (annotation and labeling), which is conducted manually or semi-automatically by annotating specific vocalizations or just the presence of a species. There is a wealth of knowledge on guidelines and challenges on the topic (e.g., Zwerts et al., 2021; Stowell, 2022) of how classification is sorted and is critical for training machine learning models to properly detect species and increase the accuracy of the overall analysis and others.

The reduced type of data analysis uses a subsampling of all recordings and only targets focal species (see below for the selection criteria of those species). For this approach, raw audio recordings are preprocessed and segmented recordings into smaller units, as above. The main difference lies in the fact that annotation and labeling are typically conducted manually and restricted to focal species vocalizations, including calls and songs. This restricted analysis can usually increase the number of potential annotators, as the level of expertise required can be lower. Automated detections can also be used since focusing on training the algorithm to detect a handful of focal species can usually be done a lot more effectively.

The abovementioned reduced analysis includes manual annotation and identification of bird detections. For this analysis, 5% of all recordings in the same type of habitat for annotation are selected. The recordings come from ARUs with the same recording effort, avoiding those that fail or do not properly record the entire sampling period. If recordings lasted 5 or fewer days, all of the recordings can be used, if not, select 5 days randomly. For the selected 5 days, 1 h per day is analyzed, usually chosen to match the duty cycle peak of the focal species of interest. For that hour, all species present are annotated. Subsequently, different time windows from that hour are selected, and the species that are present within each time window are annotated (see Table 1 for time window recommendations). The list of species detected in each time window combination is then contrasted relative to the total species that was detected during the full 1-h file. This step can help identify which time window gives more than 75% of all species detections in community detection and 100% of the focal species.

We carried out our study in a coffee landscape approximately 1,200 m elevation in the San Ramón Valley, Costa Rica, in 2021. All of the farms we selected were subject to similar incentives for regenerative agriculture and did not vary much in terms of elevation and other confounding factors. We selected 40 farms in this landscape using a phased stratified sampling design largely based on farm size (farm sizes are between 1 and 12.6 ha.), aboveground biomass at 30 m resolution (Baccini et al., 2012), and percentage of forest in the surrounding landscape with a resolution diameter of 25 m and an accuracy of 2.3 cm (Duncanson et al., 2020). We deployed one ARU per coffee plot per farm for all 40 farms during two 40+ day periods: 45 days between January and March during the pruning season and 41 days between July and August 2021 during the agrochemical application season.

We selected one plot per farm since plots were management units within a farm and were physically divided by living fences and subjected to different agricultural management practices. A total of 40 ARUs were installed 3 m aboveground on tall trees in the middle of each plot, surrounded by a homogeneous coffee habitat (e.g., shade coffee, sun coffee, or organic coffee) of a radius of at least 50 m. To avoid border effects and disrupting noise being detected in the recorders, each plot selected was at least 200 m away from a main road and 100 m away from a secondary road. The trees in which the recorders were installed were not subject to any management practice. Each ARU was programmed to record during the morning chorus, 30 min before dawn and 3.5 h after, for a total of 4 h/day (from 4:30 to 8:30 h). To register bird vocalizations during seasons when different agricultural practices were applied, the ARUs were set to record for 45 days between January (pruning season) and March and 45 days between July and August 2021 (agrochemical application season). Farmers were asked to annotate the dates when each of the agricultural practices (e.g., pruning of coffee trees and application of agrochemicals) was conducted. The types of agrochemicals applied can be found in Table 3 .

We developed and used the R package named soxr (Vega-Hidalgo, 2024) to compress the audio files to FLAC. We used soxr because it is executable on a per-deployment basis, instead of sox-o-matic that can only be used on a per-recorder basis (Center for Conservation Bioacoustics, Cornell Lab of Ornithology 2024). These tools not only assist in the compression process but also in the renaming of files (For more information on the storage method, see the complementary material).

Approximately 6 focal species were selected for this study, out of 40+ species detected in the base recordings, to represent the effects of agricultural management practices on avian habitat use based on the following: 1) species frequently detected in coffee plots on previous surveys that represent different foraging guilds or foraging strategies (nectarivores, granivores, insectivores, understory gleaners, and soil scratchers), 2) species detected in the recordings vocalizing within the coffee plots and have distinctive calls and songs, and 3) resident species that were going to be present year-round during the execution of the different agricultural management practices. For this case study, we selected six species that met the criteria outlined above. The selected species were as follows: 1) rufous-tailed hummingbird (Amazilia tzacatl), a nectarivore which is a current nester on coffee plants; 2) chestnut-capped warbler (Basileuterus delatrii), an insectivore known for eating the coffee berry borer beetle and frequently detected on coffee farms (Chain-Guadarrama et al., 2019); 3) white-tipped dove (Leptotila verreauxi), a species that forages by scratching the ground surface between coffee rows; 4) rufous-breasted wren (Pheugopedius rutilus), an insectivore and gleaner that is not as common; 5) white-eared ground sparrow (Melozone leucotis), a species that feeds mostly of seeds on the ground and is territorial and commonly found in thicket habitats (Stiles et al., 1989); and 6) Cabanis’s wren (Cantorchilus modestus), a common insectivore in coffee farms.

We used the Raven Pro 1.6 sound analysis software (Cornell University Laboratory of Ornithology) to annotate each time a species was detected on a 20-min sample within each hour of recording per day (reduced analysis approach). We used the unique six-digit species code for the annotations. Only the vocalizations evident in the spectrogram were included; signals that could be barely heard in the background were excluded. If the vocalization recurred at intervals shorter than 5 s, the entire sequence of calls was recorded within a single annotation box, capturing all occurrences of the vocalization. If the vocalization could not be effectively identified as one of the focal species but is suspected to be, a question mark was placed instead. All annotators used the same Raven presets to generate and navigate spectrograms.

Additionally, we designed a mobile application called Euphonia (see Supplementary Material ). This approach also allows for more than one person to collaborate on annotating a recording at a time and shows promise in engaging with local stakeholders in the annotation process. The app allows for the annotation of recordings using a mobile device, and the same presets are available anywhere you can access an internet connection. The addition of having focal species as “presets” was time-efficient since it did not require the user to enter the species code manually when saving each annotation box.

To determine if there was a significant difference in habitat use by the target bird species before and after each management practice (e.g., comparison of two proportions or A/B test), we used a Bayesian test of proportions using Bayesian First Aid and Rjags packages in the R Statistical Computing Environment (R Core Team, 2023). The test of proportions is a robust approach that allows us to test for differences in habitat use, even when there are differences in sampling effort (Keysers et al., 2020). The test of proportions estimates the probability of an event (e.g., posterior mean) and 95% credibility interval. To run the test of proportions, we created a species-specific matrix of detection and non-detection events before (A) and after (B) each agricultural practice. We then estimated the posterior mean of the probability of detecting a species (Θ) in a coffee plot before (Θ1) or after (Θ2) the management intervention. We generated 5,000 MCMC samples for each test and presented the posterior mean and 95% credibility interval. We also estimated the difference between the posterior means (Θ1 − Θ2) to estimate the magnitude of the change in the probability of detecting a species, and the % value is the proportion of posterior estimates that are higher in a coffee plot before (posterior distribution of Θ1) the management practice (posterior distribution of Θ2), which can be interpreted as the probability that the management practice negatively influenced habitat use by a species, expressed as a percentage.

We were able to evaluate the effect of two key coffee agricultural management practices on biodiversity—pruning and the application of agrochemicals, through the use of ARUs and using focal bird species as indicators. The six focal species showed different responses to both practices inside coffee farms, representing the complexity of bird community responses to coffee production. White-eared ground sparrow and chestnut-capped warbler are species that forage within the coffee plots (Stiles et al., 1989; Sandoval and Mennill, 2012) and are the two species that showed the greatest decrease in estimates of probability of habitat use (Θ) after the application of agrochemicals. Chestnut-capped warbler forages on the coffee plant branches looking for insects (Stiles et al., 1989); meanwhile, white-eared ground sparrow feeds mostly on seeds and small insects on the ground (Stiles et al., 1989; Sandoval and Mennill, 2012; Juárez et al., 2020). Our results indicate that habitat use by both species is likely to significantly decrease after the application of agrochemicals and is likely driven by the decline of insect abundance after the application of insecticides (BioQuim, 2015) and other chemicals ( Table 3 ) that are toxic and irritating to humans and other animals.

White-eared ground sparrows and white-tipped doves showed a decrease in the estimates of probability of habitat use (Θ) after pruning. Vegetation density is often reduced in coffee plots after pruning, and most leaves and branches are commonly left on the ground. Given that both species forage mostly on the ground looking for seeds (Stiles et al., 1989), this practice increases the obstacles for foraging on the ground and opens the vegetation coverage, exposing these species to a more open area, leading birds to search for food in other sites. However, for the chestnut-capped warbler, probability (Θ) values are higher after pruning. This warbler often forages for insects and perches within the coffee branches; hence, the elimination of coffee branches may cause a disturbance on its foraging behavior indirectly affecting its detection on the recorders.

The rufous-breasted wren, which prefers tangled vines and scrubs at the coffee plot edges, showed similar estimates of habitat use before and after pruning but decreased after the application of agrochemicals. Conversely, Cabanis’s wren, a species that moves between the coffee plants and the surrounding bushes, live fences, and riparian forests (Stiles et al., 1989), showed similar detection before and after the application of agrochemicals and pruning. Rufous-tailed hummingbirds showed similar probability estimates (Θ) before and after pruning and a slight decrease after agrochemical applications, but it was not conclusive. This lack of an effect might be driven by a reduction in the ability to detect this hummingbird due to the territorial behavior around flower patches (Stiles et al., 1989). Lower patches are not that common within the coffee plots, and the blooming of coffee plants in this area is between April and June (a period not included in the recordings). In summary, our results showed that both agricultural management practices impact the ability of coffee farms to serve as supplemental habitats for the focal bird species and likely the avian community as a whole.

The comparison of the probability of each event aided in communicating these results with coffee producers and other stakeholders, as evidence of the effect that different agricultural management practices have on biodiversity in coffee farms. This is crucial especially for the chemicals reported ( Table 3 ) such as the fungicide epoxiconazole, the use of which has been debated in Europe since 1990s, due to its high toxicity in water environments and hazardous effects on humans (European Commission, 2022; Cocco, 2002) and other vertebrates (Zhang et al., 2018), and it is used in Costa Rican coffee farms under government approval (Barquero, 2013; Ministerio de Agricultura y Ganadería, 2022).

Based on our results on six species that represent foraging guilds of common birds in coffee farms, we provided the following management suggestions: 1) consider using selective pruning or systematic cyclical pruning to maintain the structural complexity of the farm relative to pruning the entire coffee plot or farm (Azhar et al., 2013); 2) consider pruning coffee farms before April or after June, to avoid the peak of the breeding season for many resident bird species (Skutch, 1950), including the breeding peak for the rufous-tailed hummingbird (Lindell and Smith, 2003); 3) refrain from using herbicides to control weeds on access trails and trails running between coffee plots and use mowing as an alternative; and 4) use fungicides and fertilizers approved for sustainable practices and in the recommended amounts (Fuhrimann et al., 2019).

Our use of the Euphonia mobile application shows promise in the potential leveraging of technology for community-driven bioacoustic annotations. This approach shows promise in leveraging local expertise with passive acoustic monitoring (PAM), thus deepening our understanding of biodiversity. This process of engaging with local communities could also help with establishing collective conservation goals and motivate community members to engage in bioacoustic annotation, a step that is especially important in tropical regions where species lists are not yet comprehensively supported by AI models for bioacoustics species identification (Sun et al., 2022). The invaluable local knowledge in species identification, when potentially incentivized, has been shown to serve as a cornerstone of conservation efforts (McElwee et al., 2020).

More importantly, integrating communities into conservation planning promotes the dissemination of conservation ethics and trust in its intrinsic value (Cetas and Yasué, 2017; Armitage et al., 2020). While AI-assisted species identification in apps like Merlin and BirdNET, developed by the Cornell Lab of Ornithology, represents significant progress by facilitating people’s interaction with the sounds of birds that surround them, these tools are currently primarily outreach and learning tools that help observers identify species in real time. However, mobilizing collective action through shared local bioacoustic species identification and annotation via a mobile app represents an exciting and novel path. Quantifying the potential efficiency boost for ARU projects and evaluating its usability by non-scientific audiences will provide guidance for the development toward an increase in inclusivity and efficiency in bioacoustics projects.