A global assessment of BirdNET performance: differences among continents, biomes, and species

A global assessment of BirdNET performance: differences among continents, biomes, and species | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article A global assessment of BirdNET performance: differences among continents, biomes, and species David Funosas, Esther Sebastián-González, Jon Morant, Oscar H. Marín Gómez, and 96 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7832874/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Recent advances in machine learning have accelerated automated species detection across diverse ecological domains, enabling large-scale, non-invasive monitoring of biodiversity. In ornithological research, coupling passive acoustic monitoring (PAM) with rapidly-developing novel identification tools such as BirdNET—a deep learning–based sound recognition algorithm—offers new opportunities for surveying vocally active bird communities. Yet, BirdNET performance across diverse ecological and biogeographic contexts remains to be quantified. Here, we present the first worldwide evaluation of BirdNET using 4,224 one-minute soundscapes from 67 sites across 28 administrative regions annotated by local experts that included 1,020 species. More specifically, we assessed the capacity of BirdNET to correctly identify individual vocalisations and characterise bird communities based on the automated analysis of passively collected soundscapes. We further analysed how its performance varies across continents, biomes, species, and minimum confidence thresholds. The proportion of correct BirdNET predictions (precision) was generally high and consistent across continents (range: 0.57–0.71 at the vocalisation level) and biomes (range: 0.55–0.76 at the vocalisation level). In contrast, the proportion of vocalisations or species successfully detected (recall) was generally lower and more heterogeneous across continents (range: 0.24–0.52 at the vocalisation level) and biomes (range: 0.34–0.72 at the vocalisation level), reflecting differences in species coverage and local ecological context. BirdNET predictive power, as measured by the Precision-Recall Area Under the Curve (PR AUC), was highest in North America, Oceania, and Europe (range: 0.16–0.23 at the vocalisation level), moderate in Central/South America (0.13), and lowest in Africa and Asia (range: 0.03–0.04). Species-specific analyses revealed substantial heterogeneity in detection accuracy, with optimal confidence thresholds varying widely by species and analytical goal. Our results establish a global reference point for BirdNET reliability and highlight where algorithmic refinement and expanded acoustic sampling are most needed. Artificial Intelligence and Machine Learning passive acoustic monitoring bird communities BirdNET deep learning automated detection confidence threshold Figures Figure 1 Figure 2 Figure 3 Figure 4 1. Introduction In recent decades, various automated and non-invasive approaches have become standard in biodiversity monitoring (Lahoz-Monfort and Magrath 2021 ). Among these, passive acoustic monitoring (PAM) has proven particularly effective for surveying diverse taxa, including anurans, bats, birds, cetaceans, and soniferous insects (Sugai et al. 2019 , Hoefer et al. 2023 , Darras et al. 2025 ). Among terrestrial taxa, birds have been primary targets of PAM (Shonfield and Bayne 2017 , Sugai et al. 2019 ), and soundscape ecology has historically drawn on acoustic ornithology (Gasc et al. 2017 ). The documented capacity of this approach to characterise avian communities and estimate population densities from passively collected soundscapes (Darras et al. 2018, Pérez-Granados and Traba 2021 ) has prompted the development of dedicated methods and analytical pipelines for automated or semiautomated bird monitoring. In recent years, the development of effective low-cost audio recorders, such as the AudioMoth (Hill et al. 2018 ) and Song Meter Micro (Wildlife Acoustics) devices, along with essential advances in automated signal recognition software, have significantly expanded the use of PAM in biodiversity research. While PAM offers numerous possibilities, it also presents challenges. One of its primary strengths—the ability to easily scale acoustic monitoring both spatially and temporally—also results in acoustic datasets far too large for manual analysis. To address this, most current projects rely on deep learning (DL) algorithms for the automated analysis of passively collected data (Stowell 2022 , Xie et al. 2022). Unfortunately, many state-of-the-art DL algorithms remain largely inaccessible to ecologists, land managers, and non-specialists lacking computational training. To bridge this gap, a new generation of user-friendly, ready-to-use sound recognition tools has emerged, helping to further streamline PAM workflows. Notably, many of these applications focus on birds as a model group, including BirdNET (Kahl et al. 2021 ), Perch (Ghani et al. 2023 ), HawkEars (Huss et al. 2025), Nighthawk (Van Doren et al. 2024 ), Chirpity (Kirkland 2024 ), and the British Trust for Ornithology’s Acoustic Pipeline (BTO 2023). Among them, BirdNET stands out for its broad taxonomic and geographic coverage, as well as its high predictive accuracy (Kahl et al. 2021 , Pérez-Granados 2023 ). BirdNET is a free bird vocalisation recognition software based on a convolutional neural network (Kahl et al. 2021 ). Its latest version (v.2.4) can identify more than 6,000 bird species worldwide, as well as a smaller set of mammals and anurans (Wood et al. 2023a , 2023b , Pérez-Granados et al. 2023 , Bota et al., 2024 ). The model analyses recordings in 3-second windows, predicting zero, one, or multiple species per segment. Each prediction is assigned a confidence score ranging from 0.01 (very low model certainty in the prediction) to 1 (very high certainty), enabling users to filter results by using a customisable minimum confidence threshold. Low confidence thresholds favor high detection rates but increase the risk of false positives, while high confidence thresholds reduce errors at the cost of missed detections (Wood and Kahl 2024 ). Two additional parameters also shape BirdNET performance: Overlap (ranging from 0 to 3 seconds), which determines the degree of temporal overlap between consecutive 3-second windows, and Sensitivity (ranging from 0.5 to 1.5), which adjusts how confidence scores are distributed across predictions. Lower Sensitivity values increase model certainty in its top predictions, while higher values make confidence scores more uniform across predictions (Pérez-Granados et al. 2025a ). Recent studies have explored how these parameters influence BirdNET performance, with the optimal configuration differing among species, regions, and research goals (Funosas et al. 2024 , Pérez-Granados et al. 2025a ). In recent years, multiple studies have used BirdNET to automatically classify bird vocalisations and derive ecological or conservation insights from the data collected (e.g. Funosas et al., 2024 , wa Maina and Njoroge 2025 , Winiarska et al. 2025 ). Recent research suggests that BirdNET can provide a reliable characterisation of bird communities in Europe, often yielding higher species richness estimates than traditional on-site surveys due to its ability to detect nocturnal and cryptic species (Funosas et al. 2024 , Winiarska et al. 2025 ). However, most prior studies evaluating BirdNET have been limited in scope, focusing either on individual species at local scales (Manzano-Rubio et al. 2022 ) or on datasets restricted to Europe and North America (Pérez-Granados 2023 ). Only a handful of studies have evaluated BirdNET performance in other regions (Amorós-Ausina et al. 2024 , Pérez-Granados and Schuchmann 2025). This geographic bias stems in part from the initial restriction of the BirdNET training set to European and North American species (Kahl et al. 2021 ). Given the recent expansion of BirdNET species coverage and its increasing adoption worldwide, there is a growing need to rigorously evaluate its performance across diverse ecological, acoustic, and biogeographic contexts. The main goal of this study is to provide the first global evaluation of BirdNET performance. Using 4,224 one-minute soundscapes recorded at 67 sites across 28 administrative regions worldwide, we aim to assess how accurately BirdNET identifies bird vocalisations and how effectively it characterises bird communities from passively collected recordings. Our study has three objectives: (1) to examine geographic and biome-level variation in BirdNET capacity to identify individual vocalisations and characterise bird communities, (2) to provide species-level performance metrics by reporting mean precision (the proportion of correct predictions) and recall (the proportion of vocalisations successfully detected) values for each of the 1,102 bird species included in the study, and (3) to analyse how BirdNET performance varies across confidence thresholds. Given the rapidly growing use of BirdNET for automated bird monitoring, we hope that our findings will help identify its strengths and limitations across diverse ecological contexts, thus guiding future use of the software and contributing to its continued improvement. 2. Methods 2.1. Acoustic data The recordings used in this study were sourced from the World Annotated Bird Acoustic Dataset (WABAD, version 2.0, Pérez-Granados et al. 2025b). WABAD comprises expert-annotated recordings collected from 72 recording sites across 28 administrative regions (mainly representing countries). To facilitate further comparisons across datasets and ensure methodological consistency, we focused the study on a subset of 4,224 one-minute soundscapes from 67 distinct recording sites (henceforth referred to as datasets) in all 28 administrative regions, having retained only those with “strong labels” (i.e. providing the precise start and end times of each bird vocalisation). The spatial distribution of the acoustic datasets used is shown in Figure 1, and metadata for each site (minutes annotated, recording device, sampling frequency, geographic region, biome, and coordinates) are provided in Supplementary Table S1. Further details about the datasets analysed in the study are available in Pérez-Granados et al. (2025c). Data coverage in this study varied substantially across geographic regions, generally aligning with continental boundaries (Figure 1). The only exception is the Americas, which we divided into North America (Canada, USA, and Mexico) and Central/South America. This division allows us to evaluate whether BirdNET performance in North America—where public training data are more abundant—differs from that in the rest of the continent. For simplicity, we refer to geographic regions as “continents” hereafter. In Europe and Central/South America we gathered data from ≥20 datasets comprising >1,100 recordings and >27,000 annotated vocalisations each. In contrast, the other continents have far fewer annotated datasets (range 3–9), recordings (range 181–498), and annotated vocalisations (>6,000–8,000; Figure 1). BirdNET species coverage, estimated as the proportion of bird species annotated by humans that was included in the algorithm, was near-complete in Europe and the Americas (≥99%). Still, substantial gaps remained in Asia (95%), Africa (84%), and especially Oceania (72%). Table 1 summarises, for each continent, 1) the number of species and vocalisations annotated, and 2) the percentage of annotated species included in the BirdNET algorithm. The ecological distribution of the recordings annotated was also uneven across biomes, categorised as in Olson et al. (2001). Tropical/subtropical and temperate broadleaf forests, as well as Mediterranean forests & shrublands, have 9,000–35,000 vocalisations annotated. Boreal and temperate coniferous forests, tropical/subtropical grasslands, and wetlands are moderately represented (>3,000–6,000 vocalisations annotated), while deserts & xeric shrublands, as well as temperate and montane grasslands & savannahs, have >500–2,000 annotated vocalisations each. BirdNET species coverage was near-complete (≥98%) in all biomes except for tropical/subtropical broadleaf forests (91%). Table 2 summarises, for each biome, 1) the number of annotated species and vocalisations, and 2) the percentage of annotated species included in the BirdNET algorithm. 2.2. Annotation procedure Expert ornithologists with in-depth knowledge of the local bird communities annotated all recordings. Audio spectrograms were analysed in Raven Pro (v1.6), with experts being allowed to adjust the software parameters at their convenience and listen and visualise the recordings as many times as needed. Each vocalisation was labeled at the species level following the Clements Checklist nomenclature (Clements et al. 2021), the taxonomy used in BirdNET-Analyzer v.2.4 (Kahl et al. 2021), thus ensuring direct comparability between human annotations and BirdNET output. Each recording site was annotated by a single observer, who delineated vocalisations using bounding boxes encompassing their temporal and frequency ranges and exported the annotations as .txt files named after the corresponding recordings. Multiple vocalisations of the same species were grouped into a single annotation box if they occurred within one second of each other; otherwise, separate annotations were created. A more thorough description of the annotation workflow is available in Pérez-Granados et al. (2025b), and both the recordings and annotations used in this study can be downloaded from Pérez-Granados et al. (2025c). 2.3. Audio analysis We processed the acoustic recordings using BirdNET-Analyzer v2.2.0 (model v2.4: BirdNET_GLOBAL_6K_V2.4_Model_FP32.tflite) via a Linux shell script interfaced with the algorithm’s Python backend, following the approach described in Funosas et al. (2024). BirdNET includes four adjustable settings that influence detection performance: Minimum occurrence frequency threshold , Confidence threshold , Overlap , and Sensitivity (for detailed descriptions of these settings and their impact, see Kahl et al. 2021, Funosas et al. 2024, Wood and Kahl 2024, or Pérez-Granados et al. 2025c). Based on prior research assessing nine combinations of Overlap and Sensitivity values (Pérez-Granados et al. 2025c), we selected an Overlap of 2 seconds, maximising performance at both vocalisation and dataset levels, and a Sensitivity of 1 (default value). Because the effects of Sensitivity on BirdNET performance vary with continent and analytical goal (e.g. identifying individual vocalisations versus characterising bird communities), we opted for the default, balanced setting. The Minimum occurrence frequency threshold defines the lowest regional and temporal occurrence frequency a species must have to be included in the list of potentially detectable species generated by BirdNET (range: 0.01–0.99). BirdNET-Analyzer v2.4 uses eBird checklist frequency data to estimate species ranges and probabilities of occurrence given geographic coordinates and week of the year (see https://github.com/birdnet-team/BirdNET-Analyzer/discussions/234). A low threshold broadens the list of potentially detectable species by including those with low likelihoods of occurrence, whereas a higher threshold restricts the list to species with the highest expected occurrence based on eBird data. We used a Minimum occurrence frequency threshold of 0.02—i.e. requiring a 2% probability of species presence for inclusion in BirdNET-generated species lists—following Funosas et al. (2024), and retained the default Confidence threshold of 0.1 to minimise the risk of missed detections and to evaluate how detection performance varies across different thresholds. 2.4. BirdNET performance assessment We evaluated BirdNET performance by comparing its predictions to expert annotations using a suite of custom R scripts (v4.2.2; R Core Team 2025) adapted, and available, from Funosas et al. (2024). Performance was assessed at two hierarchical levels: 1) the vocalisation level, providing detailed insight into BirdNET capacity to correctly identify individual songs or calls, and 2) the dataset level, reflecting its capacity to characterise bird community composition based on multiple recordings from a single recording site. BirdNET predictions were classified into four categories: True Positives (TP): At the vocalisation level, a BirdNET prediction was classified as a TP when an expert labelled the same species at the same 3-second time interval. At the dataset level, a bird species was considered a TP if there was at least one correct identification of that species by BirdNET in any of the recordings from the same study site (i.e. dataset). False Positives (FP): At the vocalisation level, a BirdNET prediction was classified as a FP when an expert did not detect the same species at the same time. At the dataset level, a bird species was considered a FP when all BirdNET predictions of that species in the dataset were incorrect. True Negatives (TN): A species was classified as a TN when it was covered by BirdNET but was neither identified by the expert nor predicted by the algorithm. False Negatives (FN): A species was classified as a FN when it was identified by the expert but not predicted by BirdNET. Based on these categorisations, we evaluated BirdNET precision, recall and False Positive Rate (FPR) at both levels of analysis. Precision quantifies the proportion of correct identifications among all BirdNET predictions, whereas recall measures the proportion of expert-identified vocalisations or species that were correctly detected by BirdNET (Pérez-Granados 2023). FPR complements these metrics by estimating the probability that BirdNET falsely detects a species absent from the acoustic sample—defined as a 3-second prediction window at the vocalisation level and as the entire dataset at the dataset level—and is computed as the number of spurious detections divided by the number of species absent from the sample but covered by BirdNET. Importantly, recall calculations include species not covered by BirdNET: any annotated species that went undetected was counted as a FN, whether because BirdNET failed to predict it despite coverage or because the species was outside the taxonomic scope of the algorithm. This choice was deliberate, as our concern is not BirdNET recall conditional on its species list, but the practical outcome for ecological applications—i.e. if BirdNET is deployed in a region, what fraction of vocalisations or species it will detect successfully. Precision, recall and FPR were calculated across 90 confidence thresholds ranging from 0.10 to 0.99 in 0.01 increments, following Funosas et al. (2024). At the vocalisation level, recall was computed by aggregating all BirdNET predictions that overlapped a given vocalisation (i.e. those starting before its end and ending after its onset). Precision was calculated by pooling all expert annotations that overlapped each prediction. At the dataset level, comparisons were based on species lists: a species was marked as correctly predicted if it appeared in both BirdNET predictions and expert annotations, as long as they both coincided in time at some point. Note that under this criterion, a single correct detection of a species suffices for it to be classified as a TP at the dataset level, thus favouring higher recall values in longer datasets. The formulas used were the following: Precision = TP / (TP + FP) Recall = TP / (TP + FN) FPR = FP / (FP + TN) To visually represent the variation of these metrics across confidence thresholds, we used the Precision-Recall (PR) curve, accompanied by its corresponding Area Under the Curve (AUC; Davis and Goadrich, 2006). PR curves plot precision against recall across all confidence thresholds, capturing the trade-off between these two metrics, with higher AUC values (range 0–1) being indicative of higher predictive power. Because PR AUC integrates precision across the entire recall range, broader recall ranges—even those including lower recall values—can yield higher AUCs. Thus, to enable fair comparisons across continents with varying recall ranges, we adjusted PR AUC scores to account for recall ranges using the following formula: Finally, we computed the F-score, which integrates precision and recall into a single performance metric: An F-score with β = 1 gives equal weight to precision and recall, whereas β > 1 emphasises recall and β < 1 emphasises precision. We computed F-scores using three β values: β = 1 as a standard value to enable comparison with previous studies, β = 0.25 to prioritise precision over recall, and β = 4 to prioritise recall over precision. A β value < 1 was chosen because high precision is typically paramount in biodiversity research, where failing to detect a present species is generally less problematic than falsely detecting an absent one (Tolkova et al. 2021). Furthermore, specific population models (e.g. occupancy models) explicitly account for imperfect detection (Brunk et al. 2023, Bielski et al. 2024), further reducing the relative cost of FNs compared with FPs. A β value > 1, on the other hand, might be particularly relevant to research teams targeting rare or cryptic species and able to manually validate large numbers of BirdNET predictions. Moreover, some classification–occupancy models can explicitly incorporate classification errors, making FPs less problematic (Ogawa et al 2025). The metrics described above were used to capture complementary aspects of BirdNET performance across continents, biomes, and species. PR AUC scores provided a global measure of predictive power, with PR curves offering a fine-grained view of how precision and recall shift with confidence threshold choice in each continent. F-score curves were used to identify, for every continent, the threshold that maximizes the trade-off between precision and recall under three weighting schemes: equal weighting, marked emphasis on precision, and marked emphasis on recall. To enable consistent cross-continent and cross-biome comparisons for different purposes—whether identifying individual vocalisations (vocalisation level) or characterising bird communities (dataset level)—, we calculated recall and precision at both levels of analysis using BirdNET’s default confidence threshold of 0.1. Two additional thresholds (0.5 and 0.75) were also used to evaluate how optimal threshold choice varies across species. We quantified uncertainty in continent- and biome-specific performance metrics using bootstrapped 95% confidence intervals across datasets within each continent and biome. 3. RESULTS 3.1. Performance across continents We found BirdNET performance to be moderately heterogeneous across continents. When using the default confidence threshold of 0.1, precision was relatively consistent across continents at the vocalisation level (range 0.57–0.71, sd = 0.053), and more variable at the dataset level (range 0.27–0.62, sd = 0.134; Figure 2), with Africa performing worst at both levels of analysis and Oceania performing best at the dataset level and second-to-best at the vocalisation level (Table 1). Recall, in turn, varied more widely across continents at the vocalisation level (range 0.24–0.52, sd = 0.125) but was more uniform at the dataset level (range 0.49–0.70, sd = 0.084), with Europe and North America performing consistently better and Africa and Asia performing consistently worse at both levels (Table 1, Figure 2, Supplementary Figure S1). Finally, FPR varied widely across continents at the vocalisation level (range 3e-05–6e-05, sd = 1.03e-05) and even more so at the dataset level (range 0.001–0.011, sd = 0.0034; Table 1), with high discrepancies in continent ranking between the two levels of analysis. Central and South America performed best and Europe performed worst at the vocalisation level, while Oceania performed best and Africa performed worst at the dataset level. At the vocalisation level, adjusted PR AUC values highlight more evident differences: Asia (0.03) and Africa (0.04) exhibit very low scores at the vocalisation level, with Central/South America (0.13) and Europe (0.16) having intermediate scores, and Oceania (0.20) and North America (0.23) performing best. Asia shows the lowest adjusted PR AUC score because, even though the precision scores obtained with high confidence thresholds are very high, these drop precipitously when the threshold is lowered (Figure 3E). At the same time, the maximum recall score for this continent is already low (0.26), implying that any confidence threshold yielding reasonable precision will necessarily drive recall to very low levels. In Africa, both maximum precision and recall are lower than in Asia, but its precision demonstrates greater robustness, decreasing more slowly with reduced confidence thresholds (Figure 3A). At the dataset level, adjusted PR AUC values follow similar trends, but with a lower degree of cross-continent heterogeneity. Africa exhibits the lowest score (0.16), while Europe, Asia, Central/South America, and North America have intermediate scores (range 0.22–0.29) and Oceania stands out (0.35) due to consistently high precision (≥0.62) across confidence thresholds (Figure 3). F-score curves show broadly similar shapes across continents. F1-scores peak at a confidence threshold of 0.1 at the vocalisation level and plateau at a mostly flat maximum between 0.2 and 0.6 at the dataset level. F0.25-scores, for the most part, plateau at a relatively flat maximum between 0.3 and 0.8 at the vocalisation level and reach a peak at a confidence threshold of around 0.9 at the dataset level. F4-scores steadily decrease along with higher confidence thresholds at both levels of analysis, with declines being particularly pronounced at the dataset level (Figure S2). 3.2. Performance across biomes Consistent with continent-scale results, we found that, when using the default confidence threshold of 0.1, precision was relatively uniform across biomes at the vocalisation level (range 0.55–0.76, sd = 0.065) but more heterogeneous at the dataset level (range 0.14–0.41, sd = 0.090; Table 2). Best- and worst-performing biomes are not entirely consistent across the two levels: BirdNET performed especially poorly in wetlands at the vocalisation level and in deserts & xeric shrublands at the dataset level. In contrast, classifications in recordings from tropical/subtropical dry broadleaf forests and montane grasslands & savannahs exhibited the highest precision, maintaining strong performance at both levels of analysis. Recall varied more widely across biomes, with values of 0.39–0.72 (sd = 0.105) at the vocalisation level and 0.51–0.85 (sd = 0.127) at the dataset level (Table 2). Scores were consistently highest in deserts & xeric shrublands, whereas tropical/subtropical broadleaf forests ranked lowest at the vocalisation level and near the bottom at the dataset level, only surpassing montane grasslands & savannahs (Table 2, Figure S3). Finally, FPR varied widely across biomes at the vocalisation level (range 3e-05–9e-05, sd = 1.91e-05) and even more so at the dataset level (range 0.003–0.016, sd = 0.0038), with broad similarities in biome ranking between the two levels of analysis (Table 2). Montane grasslands & savannahs performed best at both levels, with temperate broadleaf & mixed forest and tropical/subtropical grasslands performing worst at the vocalisation and dataset levels, respectively. Confidence intervals for mean performance across both continents and biomes overlapped partially across all three metrics and both levels of analysis (Table 1), indicating moderate heterogeneity rather than clear-cut regional contrasts. Some of this overlap likely reflects uneven sample sizes across continents and biomes, as regions with fewer datasets exhibited broader confidence intervals and hence greater uncertainty in estimated means. 3.3. Performance across species Our species-specific analyses, focused on the species annotated in more than 50 recordings and conducted across three confidence thresholds (0.1, 0.5, 0.75), show that, as the minimum confidence threshold increases, both TPs and FPs decline, with FNs following the opposite trend (Figure 4). Consequently, the numbers of correctly and mistakenly detected species vary substantially with threshold choice. Because higher-confidence predictions are more likely to be correct, raising the threshold reduces FPs far more than TPs. This asymmetry, however, is highly species-dependent (Figure 4, Supplementary Table S2). Some species (e.g. Arremon brunneinucha ) retain most TPs while eliminating nearly all FPs when increasing the confidence threshold from 0.1 to 0.75, achieving consistently high F1-scores (>0.7) across thresholds. However, other species (e.g. Acrocephalus arundinaceus ) lose ≥90% of TPs with the same confidence threshold increase, maintaining persistently low F1-scores (<0.3, Figure 4). Vocalisation-level precision and recall scores, along with the numbers of species-specific annotated vocalisations and BirdNET predictions for all 1,102 species in the WABAD dataset, are reported in Supplementary Table S2. 4. DISCUSSION The use of deep learning algorithms for automated wildlife detection from passively collected data has expanded rapidly in recent years (Stowell 2022 , Xie et al. 2022). Here, we provide the first comprehensive evaluation of BirdNET (Kahl et al. 2021 ) performance across continents and biomes, assessing both its ability to detect bird vocalisations and to characterise bird communities. We analysed an extensive acoustic dataset composed of 4,224 one-minute soundscapes annotated by local experts from 67 recording sites worldwide (Pérez-Granados et al. 2025a ). Our results suggest that BirdNET cross-continent and cross-biome performances are highly heterogenous, with recall being more variable than precision. Much of this variation can be attributed to gaps in BirdNET species coverage, highlighting the importance of consulting the most up-to-date BirdNET species list when interpreting results. The fact that precision remains imperfect even at the highest confidence thresholds, and that relaxing these thresholds is unavoidable to obtain reasonable recall, makes expert validation indispensable for achieving accurate and comprehensive ecological inference. Precision exhibited substantial variation across continents and biomes, with vocalisation-level estimates being more consistent than those at the dataset level. Dataset-specific factors, including recording duration—which tends to increase the number of FPs at fixed thresholds (Funosas et al. 2024 )—and species richness—where lower values magnify the relative impact of FPs—likely contributed to this discrepancy. At the biome scale, recordings collected in wetland habitats exhibited the weakest performance at the vocalisation level. Since many wetland species are widespread and well represented in reference libraries, low performance possibly reflects ecological complexity driven by high species richness and abundance (Goëau et al. 2018 ), as well as the presence of species with relatively uncharacteristic or poorly differentiated calls (e.g. many waterfowl). Unlike passerines, where strong sexual selection and territoriality have driven the evolution of distinctive songs, waterfowl typically depend more on visual courtship displays, reducing pressure for acoustically distinctive signals (Johnsgard 1971 , Ten Cate 2021 ) and thereby complicating discrimination by BirdNET. Deserts, by contrast, exhibited the lowest dataset-level precision despite strong recall: with few species annotated per site, even a small number of incorrect detections can drastically reduce precision. At the continental scale, precision was found to be the weakest in Africa and Asia at both levels of analysis, possibly due to the relatively small amount of training data available for species occurring in these continents (Funosas et al. 2024 ). Oceania, however, achieved the highest dataset-level precision and the second-highest vocalisation-level precision despite having the lowest degree of species coverage. This result may be partly explained by the partnership established between the Cornell Lab of Ornithology, developer of BirdNET, and the Listening Observatory for Hawaiian Ecosystems, possibly leading to the algorithm being better adjusted to the Hawaiian avifauna. FPR broadly tracked precision, with vocalisation-level estimates more consistent than dataset-level ones. At the dataset level, continental rankings were nearly identical: Africa and Asia performed worst, Oceania and North America best. At the vocalisation level, however, Central and South America showed the lowest FPR, while Europe had the highest—indicating that European precision is penalised mainly by its high number of FPs, whereas Central and South American precision is mainly driven down by its low number of TPs. At the biome scale, the biomes with the highest precision (tropical/subtropical dry broadleaf forests and montane grasslands & savannahs) also had the lowest FPR, while those with the highest FPR (temperate broadleaf & mixed forests and temperate coniferous forests at the vocalisation level; tropical/subtropical and temperate grasslands at the dataset level) also exhibited the lowest precision. This tight correspondence highlights the central role of FPs in shaping both metrics. Recall also displayed broadly similar patterns to precision: continents and biomes with the most significant proportion of uncovered species (e.g. Africa, Asia, tropical/subtropical broadleaf forests) generally exhibited the lowest scores, with Oceania again standing out as an exception. Oceanian recall was surpassed only by Europe and North America and clearly outperformed Africa and Asia at both levels of analysis (Table 1 ). Beyond the potential emphasis on Hawaiian avifauna during BirdNET training, the low species richness in the geographically isolated Hawaiian bird communities—accounting for half of our Oceanian data—may also have contributed to higher detection rates by increasing the likelihood of detecting a large share of the present species. A similar pattern was observed in cross-biome analyses: deserts & xeric shrublands, which have the lowest species richness, achieved the highest recall at both levels of analysis (Table 2 ). Taken together, these results underscore the combined influence of species coverage, dataset composition, and local context on BirdNET performance. However, given the limited number of datasets and uneven spatial coverage for certain continents and biomes—most notably Oceania and deserts—, these apparent patterns may partly reflect sampling artifacts rather than generalisable performance differences. When integrating across the F1-score and PR AUC metrics, BirdNET performed best in North America, Europe, and Oceania, moderately in Central/South America, and relatively poorly in Asia and Africa, with these cross-continent patterns largely consistent between both levels of analysis. Geographic differences likely reflect disparities in training data: online acoustic libraries are richer in recordings from North America and Europe, whereas African and Asian species remain underrepresented (Macaulay 2025 , Xeno-canto 2025 ). The paucity of species-specific training data in these regions may predispose the algorithm to learn narrow or context-specific acoustic cues that fail to generalise more broadly. This could undermine the capacity of BirdNET to recognise species whose vocal signatures vary geographically or ecologically (Knight et al. 2024 , Sebastián-González and Pérez-Granados 2025 ), as well as species with very wide acoustic repertoires. Confidence threshold analyses, in contrast, highlight a high degree of consistency across continents but a large difference between scales of evaluation. At the vocalisation level, F1-scores peak at a confidence threshold of 0.1 and decline steadily at higher thresholds, as recall decreases more sharply than precision improves. F0.25-scores plateau at intermediate thresholds because the greater weighting of precision over recall means that small precision gains can compensate for moderate recall losses. F4-scores, driven mainly by recall and only weakly by precision, decline steadily with higher confidence thresholds (Figure S2). At the dataset level, since only one correct detection is required for a species to be considered a TP, higher thresholds can improve results by favoring precision without as steep a recall penalty. This may explain why both recall rates and optimum thresholds are higher at the dataset level. These differences highlight that there is no universally optimal confidence threshold: the best choice is inherently context-dependent (Wood and Kahl 2024 , Tueng et al. 2025). Lower thresholds enhance recall and are therefore better suited to the monitoring of rare or elusive species, whereas higher thresholds are preferable for biodiversity monitoring, where minimizing FPs is essential to prevent inflated richness estimates. The instability of precision-focused F-scores at high thresholds also cautions against overconfidence, highlighting the need for manual validation—especially in regions with low training coverage. More broadly, threshold selection is not just a technical choice but a reflection of ecological or conservation priorities. Explicitly reporting and justifying threshold choices could therefore enhance reproducibility and interpretability in future BirdNET applications. Our species-level analyses also revealed substantial heterogeneity in BirdNET vocalisation-level performance across species (Fig. 4 , Supplementary Table S2), a result that aligns with prior research suggesting strong variation across species, including within the same family (Amorós et al. 2024). Strong cross-context intraspecific variation in BirdNET performance has also been reported, with mean precision for the Common Raven ( Corvus corax ) ranging from 0.29 (Cole et al. 2022 ) to 0.66 (Kahl 2020 ) and 0.94 (Sethi et al. 2021 ). In our study, interspecific variability in vocalisation-level precision and recall diminished with increasing confidence thresholds, with standard deviations shrinking from 0.23 and 0.19 at a threshold of 0.1 to 0.18 and 0.13 at a threshold of 0.75. As expected, increasing thresholds improved precision but reduced recall, though the magnitude of this trade-off varied widely across species. In some cases, TPs declined far more slowly than FPs, making these species better suited to higher thresholds. In others, TPs and FPs declined at similar rates (Fig. 4 ), favoring lower thresholds. While the use of uniform thresholds for all species remains a practical option, these results underscore that species-specific adjustments are likely to yield more reliable outcomes (Wood and Kahl 2024 , Tseng et al. 2025 ). Despite having limited our species-level analysis to those present in at least 50 recordings, there was substantial heterogeneity in the number of annotated vocalisations per species. Some species (e.g. Fringilla coelebs , Turdus merula , Erithacus rubecula , Sylvia atricapilla ) had more than 1000 annotated vocalisations across over 300 recordings, while others (e.g. Herpsilochmus rufimarginatus , Curruca undata , Phasianus colchicus , Eurillas curvirostris ) had fewer than 200 annotations drawn from only 50 recordings. Results for species with sparse representation—whether in annotated vocalisations or BirdNET predictions—should therefore be interpreted with particular caution. Moreover, several species, despite being frequent in the dataset, were largely undetected by BirdNET because they were excluded from many auto-generated species lists. A notable case is Abroscopus albogularis , present in 969 recordings but included in only 111 lists and predicted in none. This result underlines the importance of revising the species lists automatically generated by BirdNET for each location and week of the year to detect potentially missing species. Since BirdNET generates these lists based on eBird data, this consideration will be especially important in undersampled regions, where lists are more likely to erroneously exclude locally present species (Table 1 , Table 2 ). Further, prior work suggests that the performance of DL acoustic classifiers is not only species- but also context-dependent, influenced by factors such as background noise and the presence of acoustically similar taxa (Ventura et al. 2024 , Tseng et al. 2025 ). This underscores the need for prudence when extrapolating our species-specific results to datasets collected under different recording settings, background noise profiles, or bird community compositions (Wood and Kahl 2024 ). Interpretation of our cross-continent and cross-biome results should also be undertaken cautiously, as the acoustic datasets analysed exhibit uneven representation across continents and biomes. Europe and Central/South America account for more than half of the data, whereas Africa, Asia, and Oceania are sparsely represented. Within continents, datasets are often concentrated in a few countries (e.g. Brazil and Colombia in Central/South America and Spain in Europe), so our findings cannot be assumed to be fully representative of entire world regions. This bias is particularly pronounced for Oceania, as our study covers only two regions in a continent with exceptional endemic diversity. Similarly, biome-level analyses reveal significant imbalances: tropical/subtropical, temperate, and Mediterranean forests are well represented, while deserts & xeric shrublands and montane grasslands & savannahs are poorly sampled in WABAD (Pérez-Granados et al. 2025c). A further limitation arises from the fact that continent and biome categories are not fully independent, since many biomes occur only in specific continents, complicating attribution. The partial overlap of cross-continent and cross-biome confidence intervals further suggests that the observed performance differences, while directionally consistent, may partly reflect these unequal sampling efforts rather than intrinsic model limitations. Consequently, apparent performance gaps should be interpreted with caution. In addition, the absence of common habitats such as croplands and urban areas, now central to biodiversity monitoring and research, limits the applicability of our findings to these environments. Additional sources of bias stem from the datasets themselves. Recording equipment and settings, annotation effort, and local conditions vary across locations, potentially influencing BirdNET performance (Leroy et al. 2021, Duc et al. 2021 , Wood and Kahl 2024 , Pérez-Granados 2025 ). Although all recordings were annotated by local experts following standardized protocols, differences in annotation quality remain possible, potentially biasing our results. Finally, we would like to acknowledge that although incorporating temporal and spatial filters (e.g., week and location) into the species list potentially detected by BirdNET is conceptually sound and improves ecological plausibility, it may inadvertently exclude species that are present but not expected according to reference databases, such as eBird. Such mismatches can affect detection performance and introduce seasonal or regional biases, affecting to a larger extent those areas with smaller reference databases. The impact of applying temporal and spatial filters on BirdNET performance should be further assessed to ensure robust and unbiased classification across contexts. Overall, our study provides a practical framework for the informed use of BirdNET in biodiversity monitoring. While precision remains relatively consistent across regions, recall drops sharply in continents and biomes with limited acoustic data availability. Responsible BirdNET use requires objective- and, ideally, species-specific confidence thresholds, coupled with manual validation to ensure a rigorous and thorough characterisation of recorded bird communities. Our findings set the first global reference for these calibration decisions and emphasise the need to fill acoustic data gaps to ensure reliable, globally consistent model performance. References Amorós-Ausina, D., Schuchmann, K. L., Marques, M. I., & Pérez-Granados, C. (2024). Living Together, Singing Together: Revealing Similar Patterns of Vocal Activity in Two Tropical Songbirds Applying BirdNET. Sensors , 24 (17), 5780. Bielski, L., Cansler, C. A., McGinn, K., Peery, M. Z., & Wood, C. M. (2024). Can the Hermit Warbler ( Setophaga occidentalis ) serve as an old-forest indicator species in the Sierra Nevada? Journal of Field Ornithology, 95 (1), 4. Bota, G., Manzano-Rubio, R., Catalán, L., Gómez-Catasús, J., & Pérez-Granados, C. (2023). Hearing to the unseen: AudioMoth and BirdNET as a cheap and easy method for monitoring cryptic bird species. Sensors , 23 (16), 7176. Bota, G., Manzano-Rubio, R., Fanlo, H., Franch, N., Brotons, L., Villero, D., ... & Pérez-Granados, C. (2024). Passive acoustic monitoring and automated detection of the American bullfrog. Biological Invasions , 26 (4), 1269–1279. Biffi, S., Chapman, P. J., Engler, J. O., Kunin, W. E., & Ziv, G. (2024). Using automated passive acoustic monitoring to measure changes in bird and bat vocal activity around hedgerows of different ages. Biological Conservation, 296, 110722. British Trust for Ornithology. (2023). BTO Acoustic Pipeline: Detection and identification of birds, bats and wildlife from audio recordings. Available at: http://bto.org/pipeline. Last accessed 15/06/2025. Brunk, K. M., Gutiérrez, R. J., Peery, M. Z., Cansler, C. A., Kahl, S., & Wood, C. M. (2023). Quail on fire: Changing fire regimes may benefit mountain quail in fire-adapted forests. Fire Ecology, 19 (1), 1–13. Clements, J. F., Schulenberg, T. S., Iliff, M. J., Billerman, S. M., Fredericks, T. A., Gerbracht, J. A., Lepage, D., Sullivan, B. L., & Wood, C. L. (2021). The eBird/Clements checklist of Birds of the World: v2021. Cole, J. S., Michel, N. L., Emerson, S. A., & Siegel, R. B. (2022). Automated bird sound classifications of long-duration recordings produce occupancy model outputs similar to manually annotated data. Ornithological Applications , 124 (2), duac003. Cord, A. F., Darras, K., Ogawa, R., Barbaro, L., Gerling, C., Kernecker, M., Markova-Nenova, N., Rodriguez-Barrera, G., Zichner, F., & Wätzold, F. (2025). Leveraging passive acoustic monitoring for result-based agri-environmental schemes: Opportunities, challenges and next steps. Biological Conservation, 305, 111042. Darras, K., Batáry, P., Furnas, B. J., Grass, I., Mulyani, Y. A., & Tscharntke, T. (2019). Autonomous sound recording outperforms human observation for sampling birds: a systematic map and user guide. Ecological Applications , 29 (6), e01954. Darras, K. F., Rountree, R., Van Wilgenburg, S., Cord, A. F., Chen, Y., Dong, L., ... & Wanger, T.C (2025). Worldwide Soundscapes: a synthesis of passive acoustic monitoring across realms. Global Ecology and Biogeography, 34(5): e70021. Duc, P. N. H., Torterotot, M., Samaran, F., White, P. R., Gérard, O., Adam, O., & Cazau, D. (2021). Assessing inter-annotator agreement from collaborative annotation campaign in marine bioacoustics. Ecological Informatics , 61, 101185. Funosas, D., Barbaro, L., Schillé, L., Elger, A., Castagneyrol, B., & Cauchoix, M. (2024). Assessing the potential of BirdNET to infer European bird communities from large-scale ecoacoustic data. Ecological Indicators , 164 , 112146. Gasc, A., Francomano, D., Dunning, J. B., & Pijanowski, B. C. (2017). Future directions for soundscape ecology: The importance of ornithological contributions. The Auk , 134 (1), 215–228. Ghani, B., Denton, T., Kahl, S., & Klinck, H. (2023). Global birdsong embeddings enable superior transfer learning for bioacoustic classification. Scientific Reports, 13(1), 22876. Goëau, H., Kahl, S., Glotin, H., Planqué, R., Vellinga, W.-P., & Joly, A. (2018). Overview of BirdCLEF 2018: Monospecies vs. soundscape bird identification. CEUR Workshops Proceedings, 2125(9). Hill, A. P., Prince, P., Piña Covarrubias, E., Doncaster, C. P., Snaddon, J. L., & Rogers, A. (2018). AudioMoth: Evaluation of a smart open acoustic device for monitoring biodiversity and the environment. Methods in Ecology and Evolution , 9 (5), 1199–1211. Johnsgard, P. A. (1971). Observations on sound production in the Anatidae. Wildfowl, 22 , 46–59. Kahl, S. (2020). Identifying Birds by Sound: Large-scale Acoustic Event Recognition for Avian Activity Monitoring. Unpublished doctoral dissertation, Chemnitz University of Technology . Kahl, S., Wood, C. M., Eibl, M., & Klinck, H. (2021). BirdNET: A deep learning solution for avian diversity monitoring. Ecological Informatics , 61 , 101236. Kirkland, M. (2024). Chirpity‑Electron: AI‑powered audio analyser for bird call visualisation, detection and cataloguing Available at: https://github.com/Mattk70/Chirpity-Electron. Last accessed: 15/06/2025. Knight, E., Rhinehart, T., de Zwaan, D. R., Weldy, M. J., Cartwright, M., Hawley, S. H., ... & Kitzes, J. (2024). Individual identification in acoustic recordings. Trends in Ecology & Evolution , 39 (10), 947–960. Hoefer, S., McKnight, D. T., Allen-Ankins, S., Nordberg, E. J., & Schwarzkopf, L. (2023). Passive acoustic monitoring in terrestrial vertebrates: a review. Bioacoustics , 32 (5), 506–531. Lahoz-Monfort, J. J., & Magrath, M. J. (2021). A comprehensive overview of technologies for species and habitat monitoring and conservation. BioScience , 71 (10), 1038–1062. Lassandro, M., Bolitho, L., & Newell, D. (2025). A systematic review of Passive Acoustic Monitoring for anuran conservation. Bioacoustics , 34(5), 579–628. Leroy, E. C., Thomisch, K., Royer, J. Y., Boebel, O., & Van Opzeeland, I. (2018). On the reliability of acoustic annotations and automatic detections of Antarctic blue whale calls under different acoustic conditions. The Journal of the Acoustical Society of America , 144 (2), 740–754. Macaulay (2025). The World’s Premier Scientific Archive of Natural History Audio, Video, and Photographs. https://www.macaulaylibrary. org/about/. Manzano-Rubio, R., Bota, G., Brotons, L., Soto-Largo, E., & Pérez-Granados, C. (2022). Low-cost open-source recorders and ready-to-use machine learning approaches provide effective monitoring of threatened species. Ecological Informatics , 72 , 101910. Ogawa, R., Gosselin, F., Darras, K. F. A., Roilo, S., & Cord, A. F. (2025). A classification-occupancy model based on automatically identified species data. Ecology, 106(5), e70086. Olson, D. M., Dinerstein, E., Wikramanayake, E. D., Burgess, N. D., Powell, G. V., Underwood, E. C., ... & Kassem, K. R. (2001). Terrestrial Ecoregions of the World: A New Map of Life on Earth: A new global map of terrestrial ecoregions provides an innovative tool for conserving biodiversity. BioScience, 51(11), 933-938. Pérez‐Granados, C. (2023). BirdNET: applications, performance, pitfalls and future opportunities. Ibis , 165 (3), 1068-1075. Pérez-Granados, C. 2025. BirdNET’s confidence scores decrease with bird distance to the recorder: revisiting Pérez-Granados (2023). Ardeola, 72 (2): 149-159. Pérez‐Granados, C., & Traba, J. (2021). Estimating bird density using passive acoustic monitoring: a review of methods and suggestions for further research. Ibis , 163 (3), 765–783. Pérez‐Granados, C., Bota, G., Giralt, D., Barrero, A., Gómez‐Catasús, J., Bustillo‐De La Rosa, D., & Traba, J. (2019). Vocal activity rate index: a useful method to infer terrestrial bird abundance with acoustic monitoring. Ibis , 161 (4), 901–907. Pérez-Granados, C., Feldman, M. J., & Mazerolle, M. J. (2023). Combining two user-friendly machine learning tools increases species detection from acoustic recordings. Canadian Journal of Zoology , 102 (4), 403–409. Pérez-Granados, C., Funosas, D., Morant, J., Gómez, O.H.M., Mendoza, I., … & Sebastián-González, E. 2025a. Optimisation of passive acoustic bird surveys: a global assessment of BirdNET settings. Preprint. Available at Research Square: DOI: 10.21203/rs.3.rs-6633549/v1. Pérez Granados, C., Morant, J., Darras, K.F.A., Gómez, O.H.M., Mendoza, I., Mohedano-Muñoz, M.A., …. & Sebastián-González, E. 2025b. WABAD: A World Annotated Bird Acoustic Dataset for passive acoustic monitoring. Preprint. Available at. Research Square: DOI: 10.21203/rs.3.rs-5729784/v1 Pérez Granados, C., Morant, J., Darras, K.F.A., Gómez, O.H.M., Mendoza, I., Mohedano-Muñoz, M.A., …. & Sebastián-González, E. 2025c. WABAD: A World Annotated Bird Acoustic Dataset for passive acoustic monitoring. Zenodo. DOI: 10.5281/zenodo.15629388 Sebastián-González, E., Camp, R. J., Tanimoto, A. M., de Oliveira, P. M., Lima, B. B., Marques, T. A., & Hart, P. J. (2018). Density estimation of sound-producing terrestrial animals using single automatic acoustic recorders and distance sampling. Avian Conservation and Ecology 13 (2), 7. Sethi, S. S., Fossøy, F., Cretois, B., & Rosten, C. M. (2021). Management relevant applications of acoustic monitoring for Norwegian nature – The Sound of Norway. NINA Report 2064. In 31. Norsk institutt for naturforskning (NINA) . Shonfield, J., & Bayne, E. M. (2017). Autonomous recording units in avian ecological research: current use and future applications. Avian Conservation & Ecology , 12 (1), 14. Stowell, D. (2022). Computational bioacoustics with deep learning: a review and roadmap. PeerJ , 10 , e13152. Sugai, L. S. M., Silva, T. S. F., Ribeiro Jr, J. W., & Llusia, D. (2019). Terrestrial passive acoustic monitoring: review and perspectives. BioScience , 69 (1), 15–25. ten Cate, C. (2021). Re-evaluating vocal production learning in non-oscine birds. Philosophical Transactions of the Royal Society B: Biological Sciences , 376(1836), 20200249. Tolkova, I., Chu, B., Hedman, M., Kahl, S., & Klinck, H. (2021). Parsing Birdsong with Deep Audio Embeddings. arXiv. Tseng, S., Hodder, D. P., & Otter, K. A. (2025). Setting BirdNET confidence thresholds: species-specific vs. universal approaches. Journal of Ornithology , 1–13. Van Doren, B. M., Farnsworth, A., Stone, K., Osterhaus, D. M., Drucker, J., & Van Horn, G. (2024). Nighthawk: acoustic monitoring of nocturnal bird migration in the Americas. Methods in Ecology and Evolution , 15 (2), 329–344. Ventura, T. M., Ganchev, T. D., Pérez-Granados, C., De Oliveira, A. G., de SG Pedroso, G., Marques, M. I., & Schuchmann, K. L. (2024). The importance of acoustic background modelling in CNN-based detection of the neotropical White-lored Spinetail (Aves, Passeriformes, Furnaridae). Bioacoustics, 33 (2), 103–121. wa Maina, C., & Njoroge, P. (2025). Comparing point counts, passive acoustic monitoring, citizen science and machine learning for bird species monitoring in the Mount Kenya ecosystem. Philosophical Transactions B , 380 (1928), 20240057. Winiarska, D., Neubauer, G., Budka, M., Szymański, P., Barczyk, J., Cholewa, M., & Osiejuk, T. S. (2025). BirdNET provides superior diversity estimates compared to observer-based surveys in long-term monitoring. Ecological Indicators , 177 , 113747. Wood, C. M., & Kahl, S. (2024). Guidelines for appropriate use of BirdNET scores and other detector outputs. Journal of Ornithology , 165: 777–782. Wood, C. M., Kahl, S., Barnes, S., Van Horne, R., & Brown, C. (2023a). Passive acoustic surveys and the BirdNET algorithm reveal detailed spatiotemporal variation in the vocal activity of two anurans. Bioacoustics , 32 (5), 532–543. Wood, C. M., Barceinas Cruz, A., & Kahl, S. (2023b). Pairing a user‐friendly machine‐learning animal sound detector with passive acoustic surveys for occupancy modeling of an endangered primate. American Journal of Primatology , 85 (8), e23507. Xie, J., Zhong, Y., Zhang, J., Liu, S., Ding, C., & Triantafyllopoulos, A. (2023). A review of automatic recognition technology for bird vocalizations in the deep learning era. Ecological Informatics , 73 , 101927. Xeno-canto (2025). Sharing Bird Sounds from Around the World. https://www.xeno-canto.org/about/xeno-canto. Tables Table 1: BirdNET performance across continents when using a minimum confidence threshold of 0.1. For each continent, the table reports the number of annotated vocalisations and species, the number and proportion of annotated species 1) correctly detected, 2) in BirdNET auto-generated lists but not detected, 3) covered but not in BirdNET auto-generated lists, and 4) not covered by BirdNET, as well as precision, recall, and FPR (reported as mean ± standard deviation across datasets, followed by the bootstrapped 95% confidence interval) calculated at both vocalisation (voc_precision, voc_recall, voc_FPR) and dataset (ds_precision, ds_recall, ds_FPR) levels. Continent Vocalisations annotated Species annotated Species correctly detected Species in auto-generated lists but not detected Species covered but not included in auto-generated lists Species not covered voc_precision ds_precision voc_recall ds_recall voc_FPR ds_FPR Africa 6455 169 82 (49%) 55 (33%) 5 (3%) 27 (16%) 0.568 ± 0.156 (0.445–0.715) 0.272 ± 0.095 (0.185–0.348) 0.236 ± 0.084 (0.156–0.299) 0.487 ± 0.079 (0.42–0.546) 4e-05 ± 3e-05 (2.04e-05–6.93e-05) 0.011 ± 0.004 (0.00706–0.0133) Asia 8410 106 62 (58%) 28 (26%) 11 (10%) 5 (5%) 0.68 ± 0.22 (0.444–0.878) 0.274 ± 0.096 (0.163–0.336) 0.257 ± 0.123 (0.183–0.338) 0.527 ± 0.151 (0.37–0.672) 5e-05 ± 2e-05 (1.84e-05–6.38e-05) 0.009 ± 0.005 (0.00323–0.0125) Central and South America 27582 482 269 (56%) 178 (37%) 28 (6%) 7 (1%) 0.707 ± 0.118 (0.654–0.758) 0.344 ± 0.132 (0.288–0.407) 0.458 ± 0.179 (0.383–0.536) 0.59 ± 0.161 (0.524–0.659) 3e-05 ± 2e-05 (2.55e-05–4.06e-05) 0.007 ± 0.005 (0.00503–0.00909) Europe 29781 182 149 (82%) 20 (11%) 11 (6%) 2 (1%) 0.605 ± 0.167 (0.543–0.669) 0.282 ± 0.1 (0.243–0.32) 0.512 ± 0.165 (0.443–0.572) 0.692 ± 0.159 (0.634–0.751) 6e-05 ± 3e-05 (4.87e-05–7.22e-05) 0.008 ± 0.003 (0.00653–0.00861) North America 8748 164 124 (76%) 37 (23%) 2 (1%) 1 (1%) 0.647 ± 0.164 (0.54–0.745) 0.371 ± 0.124 (0.297–0.446) 0.517 ± 0.255 (0.353–0.657) 0.696 ± 0.174 (0.587–0.809) 5e-05 ± 5e-05 (2.36e-05–8.49e-05) 0.005 ± 0.003 (0.00352–0.00649) Oceania 8085 54 32 (59%) 5 (9%) 2 (4%) 15 (28%) 0.692 ± 0.131 (0.572–0.78) 0.624 ± 0.23 (0.426–0.823) 0.448 ± 0.358 (0.166–0.73) 0.611 ± 0.292 (0.341–0.81) 5e-05 ± 5e-05 (1.02e-05–9.17e-05) 0.001 ± 0.001 (0.000384–0.00161) Table 2: BirdNET performance across biomes when using a minimum confidence threshold of 0.1. For each biome, the table reports the number of annotated vocalisations and species, the number and proportion of annotated species 1) correctly detected, 2) in BirdNET auto-generated lists but not detected, 3) covered by BirdNET but not included in BirdNET auto-generated lists, and 3) not covered by BirdNET, as well as precision, recall, and FPR (reported as mean ± standard deviation across datasets, followed by the bootstrapped 95% confidence interval) calculated at both vocalisation (voc_precision, voc_recall, voc_FPR) and dataset (ds_precision, ds_recall, ds_FPR) levels. Standard deviations and confidence intervals are not provided for biomes being represented by a single dataset. Biome Vocalisations annotated Species annotated Species correctly detected Species in auto-generated lists but not detected Species covered but not included in auto-generated lists Species not covered voc_precision ds_precision voc_recall ds_recall voc_FPR ds_FPR Boreal Forest/Taiga 5604 79 56 (71%) 23 (29%) 0 (0%) 0 (0%) 0.586 ± 0.191 (0.445–0.748) 0.347 ± 0.088 (0.281–0.42) 0.411 ± 0.149 (0.275–0.521) 0.636 ± 0.117 (0.545–0.73) 5e-05 ± 4e-05 (2.72e-05–8.01e-05) 0.005 ± 0.003 (0.00268–0.00767) Deserts & Xeric Shrublands 623 13 11 (85%) 2 (15%) 0 (0%) 0 (0%) 0.675 0.143 0.72 0.846 4e-05 0.01 Mediterranean Forests & Shrublands 10572 116 79 (68%) 27 (23%) 8 (7%) 2 (2%) 0.641 ± 0.138 (0.555–0.717) 0.218 ± 0.104 (0.165–0.29) 0.528 ± 0.153 (0.434–0.611) 0.633 ± 0.165 (0.536–0.729) 5e-05 ± 2e-05 (3.85e-05–5.85e-05) 0.009 ± 0.003 (0.00739–0.0103) Montane Grasslands & Savannahs 529 32 19 (59%) 12 (38%) 1 (3%) 0 (0%) 0.761 ± 0.014 (0.751–0.771) 0.357 ± 0.017 (0.345–0.368) 0.452 ± 0.086 (0.392–0.513) 0.507 ± 0.044 (0.476–0.538) 3e-05 ± 0 (3.13e-05–3.65e-05) 0.003 ± 0.001 (0.00292–0.00369) Temperate Broadleaf & Mixed Forest 10099 118 96 (81%) 21 (18%) 1 (1%) 0 (0%) 0.598 ± 0.166 (0.476–0.705) 0.373 ± 0.064 (0.331–0.419) 0.547 ± 0.269 (0.346–0.712) 0.855 ± 0.099 (0.788–0.928) 9e-05 ± 5e-05 (5.07e-05–0.000125) 0.006 ± 0.002 (0.00539–0.00758) Temperate Coniferous Forest 3611 50 43 (86%) 5 (10%) 2 (4%) 0 (0%) 0.644 ± 0.15 (0.539–0.75) 0.401 ± 0.034 (0.377–0.425) 0.657 ± 0.149 (0.552–0.762) 0.878 ± 0.093 (0.812–0.944) 8e-05 ± 6e-05 (4.18e-05–0.00012) 0.005 ± 0.002 (0.00431–0.00662) Temperate Grasslands 2107 34 24 (71%) 8 (24%) 2 (6%) 0 (0%) 0.672 0.245 0.533 0.706 6e-05 0.011 Tropical/Subtropical Grasslands 4281 91 57 (63%) 31 (34%) 1 (1%) 2 (2%) 0.676 ± 0.056 (0.636–0.716) 0.221 ± 0.005 (0.217–0.224) 0.441 ± 0.172 (0.319–0.562) 0.626 ± 0.099 (0.556–0.696) 6e-05 ± 1e-05 (5.33e-05–7.26e-05) 0.016 ± 0.002 (0.0139–0.0171) Tropical/Subtropical Dry Broadleaf Forest 9334 191 118 (62%) 50 (26%) 5 (3%) 18 (9%) 0.753 ± 0.072 (0.704–0.802) 0.415 ± 0.199 (0.291–0.556) 0.387 ± 0.269 (0.212–0.575) 0.575 ± 0.198 (0.44–0.701) 3e-05 ± 2e-05 (1.52e-05–4.03e-05) 0.004 ± 0.003 (0.00216–0.00574) Tropical/Subtropical Moist Broadleaf Forest 35687 605 331 (55%) 197 (33%) 42 (7%) 35 (6%) 0.659 ± 0.155 (0.594–0.724) 0.37 ± 0.165 (0.306–0.439) 0.409 ± 0.208 (0.324–0.496) 0.577 ± 0.169 (0.509–0.641) 4e-05 ± 3e-05 (2.66e-05–4.9e-05) 0.006 ± 0.004 (0.00484–0.00813) Wetland 6614 183 106 (58%) 72 (39%) 5 (3%) 0 (0%) 0.547 ± 0.226 (0.394–0.711) 0.271 ± 0.068 (0.227–0.318) 0.505 ± 0.14 (0.415–0.604) 0.62 ± 0.127 (0.536–0.7) 6e-05 ± 2e-05 (4.12e-05–7.34e-05) 0.009 ± 0.005 (0.00556–0.0117) Additional Declarations The authors declare no competing interests. Supplementary Files Supplementarymaterial.docx Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-7832874","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":527993366,"identity":"d1979a4c-dc41-4ff3-b5ef-90a3a980a2ce","order_by":0,"name":"David Funosas","email":"data:image/png;base64,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","orcid":"https://orcid.org/0000-0001-9964-1120","institution":"Station d’Écologie Théorique et Expérimentale","correspondingAuthor":true,"prefix":"","firstName":"David","middleName":"","lastName":"Funosas","suffix":""},{"id":528006563,"identity":"66173126-482e-46da-bffc-f24d970bec1f","order_by":1,"name":"Esther Sebastián-González","email":"","orcid":"https://orcid.org/0000-0001-7229-1845","institution":"Alicante University","correspondingAuthor":false,"prefix":"","firstName":"Esther","middleName":"","lastName":"Sebastián-González","suffix":""},{"id":528006564,"identity":"685151b4-2af0-4d8f-adf1-b764ba410ad1","order_by":2,"name":"Jon Morant","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Jon","middleName":"","lastName":"Morant","suffix":""},{"id":528006565,"identity":"473c994c-eec9-4885-b8b7-667bbc85b0d3","order_by":3,"name":"Oscar H. Marín Gómez","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Oscar","middleName":"H. Marín","lastName":"Gómez","suffix":""},{"id":528006566,"identity":"0d1aca30-c327-4f14-81ca-a8160010cd21","order_by":4,"name":"Irene Mendoza","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Irene","middleName":"","lastName":"Mendoza","suffix":""},{"id":528006567,"identity":"e9444555-5139-4e37-9378-000362fd0f1b","order_by":5,"name":"Miguel A. Mohedano-Muñoz","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Miguel","middleName":"A.","lastName":"Mohedano-Muñoz","suffix":""},{"id":528006568,"identity":"dcfb808b-9c86-4a1e-9ff5-ec4e636e4f05","order_by":6,"name":"Eduardo Santamaría","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Eduardo","middleName":"","lastName":"Santamaría","suffix":""},{"id":528006569,"identity":"ee09696e-3594-4e5e-bfcf-19ad35d26d84","order_by":7,"name":"Giulia Bastianelli","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Giulia","middleName":"","lastName":"Bastianelli","suffix":""},{"id":528006570,"identity":"8c3cc268-0719-403a-8dc9-2aa1388060b7","order_by":8,"name":"Alba Márquez-Rodríguez","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Alba","middleName":"","lastName":"Márquez-Rodríguez","suffix":""},{"id":528006571,"identity":"d166e2e0-150c-4009-a73a-4fbc12feb0aa","order_by":9,"name":"Michał Budka","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Michał","middleName":"","lastName":"Budka","suffix":""},{"id":528006572,"identity":"959ba416-4843-459a-b9b2-94c3cb630c25","order_by":10,"name":"Gerard Bota","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Gerard","middleName":"","lastName":"Bota","suffix":""},{"id":528006573,"identity":"5fccb25d-4247-4692-bc49-e1a4ea56f08b","order_by":11,"name":"José M. De la Peña-Rubio","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"José","middleName":"M. De la","lastName":"Peña-Rubio","suffix":""},{"id":528006574,"identity":"898214b8-d54a-4b27-b745-9f948c4608af","order_by":12,"name":"Eladio García de la Morena","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Eladio","middleName":"García de la","lastName":"Morena","suffix":""},{"id":528006575,"identity":"c765a90f-e590-47e5-bcff-18bc594feef4","order_by":13,"name":"Manu Santa-Cruz","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Manu","middleName":"","lastName":"Santa-Cruz","suffix":""},{"id":528006576,"identity":"a86158c7-9a83-43f0-a6c9-f1f0100ba9ce","order_by":14,"name":"Pablo de la Nava","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Pablo","middleName":"de la","lastName":"Nava","suffix":""},{"id":528006577,"identity":"ce9a07c1-a9c1-4174-8cea-f06f82494871","order_by":15,"name":"Mario Fernández-Tizón","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Mario","middleName":"","lastName":"Fernández-Tizón","suffix":""},{"id":528006578,"identity":"58976dc2-70dc-407a-9ded-06ac901e0073","order_by":16,"name":"Hugo Sánchez.Mateos","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Hugo","middleName":"","lastName":"Sánchez.Mateos","suffix":""},{"id":528006579,"identity":"c9a918b6-e730-4497-924c-9337030a4635","order_by":17,"name":"Adrián Barrero","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Adrián","middleName":"","lastName":"Barrero","suffix":""},{"id":528006580,"identity":"7eb8840b-fbd2-41c4-ad59-a1470c08ce00","order_by":18,"name":"Juan Traba","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Juan","middleName":"","lastName":"Traba","suffix":""},{"id":528006581,"identity":"bd754e39-4d8c-4da3-bfb9-fd3866435507","order_by":19,"name":"Tomasz S. Osiejuk","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Tomasz","middleName":"S.","lastName":"Osiejuk","suffix":""},{"id":528006582,"identity":"91eff471-a4ad-43d9-af88-83ba3d9377bd","order_by":20,"name":"Cristina D. Alonso-Moya","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Cristina","middleName":"D.","lastName":"Alonso-Moya","suffix":""},{"id":528006583,"identity":"0592d726-800b-40fa-a133-bdc1da704c1c","order_by":21,"name":"Patrick J. Hart","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Patrick","middleName":"J.","lastName":"Hart","suffix":""},{"id":528006584,"identity":"e9a8b60b-c9d7-4808-9f51-5b5051dc2e8e","order_by":22,"name":"Amanda K. Navine","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Amanda","middleName":"K.","lastName":"Navine","suffix":""},{"id":528006585,"identity":"4c04f6d8-7cec-4896-a7e9-068e65b9c132","order_by":23,"name":"Andrés F. Montoya Muñoz","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Andrés","middleName":"F. Montoya","lastName":"Muñoz","suffix":""},{"id":528006586,"identity":"1b9dc717-ce92-47bb-9a41-06b8887d669a","order_by":24,"name":"Carlos B. de Araujo","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Carlos","middleName":"B.","lastName":"de Araujo","suffix":""},{"id":528006587,"identity":"76b01e72-e1f6-4570-8493-a9fb7ac049bb","order_by":25,"name":"Gabriel L. M. Rosa","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Gabriel","middleName":"L. M.","lastName":"Rosa","suffix":""},{"id":528006588,"identity":"0daf8d42-12ba-4dc9-91c3-c8b789552797","order_by":26,"name":"Ingrid M. D. Torres","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Ingrid","middleName":"M. D.","lastName":"Torres","suffix":""},{"id":528006589,"identity":"8af19dc7-7474-49f9-b407-0d7a1a920766","order_by":27,"name":"Ana L. C. Catalano","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Ana","middleName":"L. C.","lastName":"Catalano","suffix":""},{"id":528006590,"identity":"753bb87d-22dd-4286-9a8d-de8a1f11e0c8","order_by":28,"name":"Cassio Rachid Simões","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Cassio","middleName":"Rachid","lastName":"Simões","suffix":""},{"id":528006591,"identity":"1492dee3-5c19-4ef9-a49f-3d6906a54526","order_by":29,"name":"Diego Llusia","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Diego","middleName":"","lastName":"Llusia","suffix":""},{"id":528006592,"identity":"8231d423-551b-414c-b49c-1609924c12b2","order_by":30,"name":"Manuel B. Morales","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Manuel","middleName":"B.","lastName":"Morales","suffix":""},{"id":528006593,"identity":"fe4cad34-3abd-416e-849a-309a037d35e4","order_by":31,"name":"Pablo Acebes","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Pablo","middleName":"","lastName":"Acebes","suffix":""},{"id":528006594,"identity":"9657eb06-3e22-4af7-8b1f-8b91305c04fc","order_by":32,"name":"Juan A. Medina","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Juan","middleName":"A.","lastName":"Medina","suffix":""},{"id":528006595,"identity":"ad819655-ba10-4222-a988-8604de41a2a5","order_by":33,"name":"Nicholas Brown","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Nicholas","middleName":"","lastName":"Brown","suffix":""},{"id":528006596,"identity":"ab0158ac-b0c5-49e1-a1fb-5067a05cad57","order_by":34,"name":"Christos Astaras","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Christos","middleName":"","lastName":"Astaras","suffix":""},{"id":528006597,"identity":"717f61d5-9a29-4979-a3c4-f521a0b9b014","order_by":35,"name":"Ilias Karmiris","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Ilias","middleName":"","lastName":"Karmiris","suffix":""},{"id":528006598,"identity":"0bfebf4c-27d7-4d38-989a-cf800fad5ae2","order_by":36,"name":"Elizabeth Navarrete","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Elizabeth","middleName":"","lastName":"Navarrete","suffix":""},{"id":528006599,"identity":"2616b706-5d38-42f5-b426-887bf3f52183","order_by":37,"name":"Maxime Cauchoix","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Maxime","middleName":"","lastName":"Cauchoix","suffix":""},{"id":528006600,"identity":"e761011c-3a91-40a5-9b08-eee475fa8106","order_by":38,"name":"Luc Barbaro","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Luc","middleName":"","lastName":"Barbaro","suffix":""},{"id":528006601,"identity":"d8007644-ca1e-47dc-824c-d0bcf945ec38","order_by":39,"name":"Dominik Arend","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Dominik","middleName":"","lastName":"Arend","suffix":""},{"id":528006602,"identity":"5053df56-6c9d-4ee2-8686-03e94a1d75d8","order_by":40,"name":"Sandra Müeller","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Sandra","middleName":"","lastName":"Müeller","suffix":""},{"id":528006603,"identity":"78a5e2c6-6f02-43bb-bb3e-d6b23d24a3c7","order_by":41,"name":"Fernando González-García","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Fernando","middleName":"","lastName":"González-García","suffix":""},{"id":528006604,"identity":"068d76e2-4b1b-4b30-a1f4-5ce70a74cd84","order_by":42,"name":"Alberto González-Romero","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Alberto","middleName":"","lastName":"González-Romero","suffix":""},{"id":528006605,"identity":"f473764d-e9c2-4ffc-98da-5b70ddacbcdc","order_by":43,"name":"Christos Mammides","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Christos","middleName":"","lastName":"Mammides","suffix":""},{"id":528006606,"identity":"6c6de9bf-ee1b-4e08-9561-f3b3d7ffd800","order_by":44,"name":"Michaelangelo Pontikis","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Michaelangelo","middleName":"","lastName":"Pontikis","suffix":""},{"id":528006607,"identity":"d036ccea-4545-4ddb-af3e-c572ac91b4dd","order_by":45,"name":"Giordano Jacuzzi","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Giordano","middleName":"","lastName":"Jacuzzi","suffix":""},{"id":528006608,"identity":"cd76436d-a14c-4ab6-93a6-f19622bb8fbb","order_by":46,"name":"Julian D. Olden","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Julian","middleName":"D.","lastName":"Olden","suffix":""},{"id":528006609,"identity":"8b145df5-b989-44bc-8150-d2276f9ebe40","order_by":47,"name":"Sara P. Bombaci","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Sara","middleName":"P.","lastName":"Bombaci","suffix":""},{"id":528006610,"identity":"37bea23b-b2e3-4122-baf5-25784644b7a4","order_by":48,"name":"Gabriel Marcacci","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Gabriel","middleName":"","lastName":"Marcacci","suffix":""},{"id":528006611,"identity":"b71ceba2-4668-47ce-9ea2-b716fae58bb8","order_by":49,"name":"Alain Jacot","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Alain","middleName":"","lastName":"Jacot","suffix":""},{"id":528006612,"identity":"b80bd231-958f-4189-961f-f294a8f645da","order_by":50,"name":"Juan P. Zurano","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Juan","middleName":"P.","lastName":"Zurano","suffix":""},{"id":528006613,"identity":"2dea7147-5935-4a65-8ac8-ad29d71cbaac","order_by":51,"name":"Elena Gangenova","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Elena","middleName":"","lastName":"Gangenova","suffix":""},{"id":528006614,"identity":"1b21c823-2c13-414d-a6a2-6aea3920a9a3","order_by":52,"name":"Diego Varela","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Diego","middleName":"","lastName":"Varela","suffix":""},{"id":528006615,"identity":"086d5aed-7214-466f-b2db-b519ddcaa939","order_by":53,"name":"Facundo Di Sallo","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Facundo","middleName":"Di","lastName":"Sallo","suffix":""},{"id":528006616,"identity":"2a8edc0c-4d0a-4ec5-840f-40181931e7ef","order_by":54,"name":"Gustavo A. Zurita","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Gustavo","middleName":"A.","lastName":"Zurita","suffix":""},{"id":528006617,"identity":"c2da498c-2ce4-453a-9862-97430161d355","order_by":55,"name":"Andrey Atemasov","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Andrey","middleName":"","lastName":"Atemasov","suffix":""},{"id":528006618,"identity":"1fa7d783-f0a6-4bd6-ae2c-fade77fb0757","order_by":56,"name":"Junior A. Tremblay","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Junior","middleName":"A.","lastName":"Tremblay","suffix":""},{"id":528006619,"identity":"224fd208-08b9-47f1-a019-11471eb0be6e","order_by":57,"name":"Vincent Lamarre","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Vincent","middleName":"","lastName":"Lamarre","suffix":""},{"id":528006620,"identity":"855a00c4-85f6-47ac-b6bc-8a068226556d","order_by":58,"name":"Anja Hutschenreiter","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Anja","middleName":"","lastName":"Hutschenreiter","suffix":""},{"id":528006621,"identity":"df9aac7a-60f0-4218-a30c-9ca562be62fa","order_by":59,"name":"Alan Monroy-Ojeda","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Alan","middleName":"","lastName":"Monroy-Ojeda","suffix":""},{"id":528006622,"identity":"8cb0f979-9505-4193-92c1-cdc06badd1e6","order_by":60,"name":"Mauricio Díaz-Vallejo","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Mauricio","middleName":"","lastName":"Díaz-Vallejo","suffix":""},{"id":528006623,"identity":"eac84513-13a5-453a-affa-2051747b90f1","order_by":61,"name":"Sergio Chaparro-Herrera","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Sergio","middleName":"","lastName":"Chaparro-Herrera","suffix":""},{"id":528006624,"identity":"66c5efcb-ccd7-4a5a-9eb3-6a8fb7e317af","order_by":62,"name":"Robert A. Briers","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Robert","middleName":"A.","lastName":"Briers","suffix":""},{"id":528006625,"identity":"984906c4-eb73-4a25-8a2b-23bdbc611624","order_by":63,"name":"Renata Sousa-Lima","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Renata","middleName":"","lastName":"Sousa-Lima","suffix":""},{"id":528006626,"identity":"682caaac-4010-4e76-8287-fa7e14f1df27","order_by":64,"name":"Thiago Pinheiro","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Thiago","middleName":"","lastName":"Pinheiro","suffix":""},{"id":528006627,"identity":"6cb685ba-4b76-4eae-b6fc-ae973db5da43","order_by":65,"name":"Wigna C. da Silva","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Wigna","middleName":"C. da","lastName":"Silva","suffix":""},{"id":528006628,"identity":"15c6c401-104b-4e36-925a-cd6ccd4475aa","order_by":66,"name":"Alice Calvente","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Alice","middleName":"","lastName":"Calvente","suffix":""},{"id":528006629,"identity":"03cc9b36-d863-4e47-aae7-4a1d526b0bba","order_by":67,"name":"Raiane V. Paz","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Raiane","middleName":"V.","lastName":"Paz","suffix":""},{"id":528006630,"identity":"9f9462b7-1c7b-42fa-8059-97c116555481","order_by":68,"name":"Carlos Salustio-Gomes","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Carlos","middleName":"","lastName":"Salustio-Gomes","suffix":""},{"id":528006631,"identity":"1026281d-4537-45f7-b3d9-8ee8f5fbb12f","order_by":69,"name":"Dorgival D. Oliveira-Júnior","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Dorgival","middleName":"D.","lastName":"Oliveira-Júnior","suffix":""},{"id":528006632,"identity":"5036031b-b619-4f1c-b86f-23243888645e","order_by":70,"name":"Cicero S. Lima-Santos","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Cicero","middleName":"S.","lastName":"Lima-Santos","suffix":""},{"id":528006633,"identity":"24a942cc-18ed-4aa4-b6d7-2860b00cd1db","order_by":71,"name":"Mauro Pichorim","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Mauro","middleName":"","lastName":"Pichorim","suffix":""},{"id":528006634,"identity":"d67388fd-c2fa-42f1-a3cc-fb5b28f2ab32","order_by":72,"name":"Anamaria Dal Molin","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Anamaria","middleName":"Dal","lastName":"Molin","suffix":""},{"id":528006635,"identity":"0d75d769-cba8-48a4-979d-634b7790ec90","order_by":73,"name":"Alexandre Antonelli","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Alexandre","middleName":"","lastName":"Antonelli","suffix":""},{"id":528006636,"identity":"7d2e87b3-1644-4836-aa52-0c5d55445551","order_by":74,"name":"Svetlana Gogoleva","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Svetlana","middleName":"","lastName":"Gogoleva","suffix":""},{"id":528006637,"identity":"36263eff-3224-4302-90b6-56aa900b1719","order_by":75,"name":"Igor Palko","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Igor","middleName":"","lastName":"Palko","suffix":""},{"id":528006638,"identity":"0aa1189f-4bb5-49b4-84a2-0c182fe0ae4a","order_by":76,"name":"Hiếu V. Trong","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Hiếu","middleName":"V.","lastName":"Trong","suffix":""},{"id":528006639,"identity":"3c4afbed-23ad-44ac-9585-276de12d072a","order_by":77,"name":"Marina H. L. Duarte","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Marina","middleName":"H. L.","lastName":"Duarte","suffix":""},{"id":528006640,"identity":"268757b8-628f-423e-8c5d-fb48a5ac1b90","order_by":78,"name":"Natalia dos Santos Saturnino","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Natalia","middleName":"dos Santos","lastName":"Saturnino","suffix":""},{"id":528006641,"identity":"eb2097e8-8f8e-493b-adcc-29f1345258d9","order_by":79,"name":"Samuel R. Silva","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Samuel","middleName":"R.","lastName":"Silva","suffix":""},{"id":528006642,"identity":"db855a26-901f-4084-ab67-1f2b1e3ab9f1","order_by":80,"name":"Ana Rainho","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Ana","middleName":"","lastName":"Rainho","suffix":""},{"id":528006643,"identity":"d699cc17-e2b2-4bec-82d2-e627fa088a26","order_by":81,"name":"Paula Lopes","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Paula","middleName":"","lastName":"Lopes","suffix":""},{"id":528006644,"identity":"db7e093d-f01c-4353-a508-2aafd9885d06","order_by":82,"name":"Karl-L. Schuchmann","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Karl-L.","middleName":"","lastName":"Schuchmann","suffix":""},{"id":528006645,"identity":"3e6fc8e6-4897-4b7f-b03c-b011e34d5c79","order_by":83,"name":"Marinêz I. Marques","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Marinêz","middleName":"I.","lastName":"Marques","suffix":""},{"id":528006646,"identity":"4a81e9b4-669c-472f-9b8e-ee7428b2c17a","order_by":84,"name":"Ana S. de Oliverira Tissiani","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Ana","middleName":"S. de Oliverira","lastName":"Tissiani","suffix":""},{"id":528006647,"identity":"6d87db19-4044-40f0-bf7f-2a7c7eddf015","order_by":85,"name":"Nick A. Littlewood","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Nick","middleName":"A.","lastName":"Littlewood","suffix":""},{"id":528006648,"identity":"57af5a26-b857-416a-826e-3b70c2b85c93","order_by":86,"name":"Mao-Ning Tuanmu","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Mao-Ning","middleName":"","lastName":"Tuanmu","suffix":""},{"id":528006649,"identity":"f72bad71-b992-47ae-be81-1c95e0ea56c2","order_by":87,"name":"Sebastian Kepfer-Rojas","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Sebastian","middleName":"","lastName":"Kepfer-Rojas","suffix":""},{"id":528006650,"identity":"26c49382-8d8f-4e27-b685-728af93ab40c","order_by":88,"name":"Andrea L. Aguilera","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Andrea","middleName":"L.","lastName":"Aguilera","suffix":""},{"id":528006651,"identity":"4ad88567-7902-4a33-95a9-4dcf61759570","order_by":89,"name":"Lluís Brotons","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Lluís","middleName":"","lastName":"Brotons","suffix":""},{"id":528006652,"identity":"27794d34-3ff5-43d4-b600-11ed6c123650","order_by":90,"name":"Mariano J. Feldman","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Mariano","middleName":"J.","lastName":"Feldman","suffix":""},{"id":528006653,"identity":"06aae6e6-6cf3-4478-8979-6bac9de2e46b","order_by":91,"name":"Louis Imbeau","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Louis","middleName":"","lastName":"Imbeau","suffix":""},{"id":528006654,"identity":"ed0db6a4-b25b-41ec-bb37-b98ee659869f","order_by":92,"name":"Pooja Panwar","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Pooja","middleName":"","lastName":"Panwar","suffix":""},{"id":528006655,"identity":"5e130adc-9c34-4e90-bbf3-edbd81fd8d81","order_by":93,"name":"Aaron S. Weed","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Aaron","middleName":"S.","lastName":"Weed","suffix":""},{"id":528006656,"identity":"52083eea-8e6e-4f5a-aee1-5b1aa78115fb","order_by":94,"name":"Anant Dehwal","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Anant","middleName":"","lastName":"Dehwal","suffix":""},{"id":528006657,"identity":"b4ca6086-da56-4278-9991-f21bd5b2962f","order_by":95,"name":"Alfredo Attisano","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Alfredo","middleName":"","lastName":"Attisano","suffix":""},{"id":528006658,"identity":"1de9bdd7-4cc4-43e7-a18c-1d2aadf62a73","order_by":96,"name":"Jörn Theuerkauf","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Jörn","middleName":"","lastName":"Theuerkauf","suffix":""},{"id":528006659,"identity":"0a23da63-c0e5-4bf1-b8ef-efa3a739dd1c","order_by":97,"name":"Eben Goodale","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Eben","middleName":"","lastName":"Goodale","suffix":""},{"id":528006660,"identity":"a37ec6cb-c51c-458d-8252-874d31c19264","order_by":98,"name":"Kevin F.A. Darras","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Kevin","middleName":"F.A.","lastName":"Darras","suffix":""},{"id":528006661,"identity":"cd35ef1f-0cd8-4899-942a-f9872db17228","order_by":99,"name":"Cristian Pérez-Granados","email":"","orcid":"https://orcid.org/0000-0003-3247-4182","institution":"Forest Science and Technology Center of Catalonia","correspondingAuthor":false,"prefix":"","firstName":"Cristian","middleName":"","lastName":"Pérez-Granados","suffix":""}],"badges":[],"createdAt":"2025-10-11 07:34:14","currentVersionCode":1,"declarations":{"humanSubjects":false,"vertebrateSubjects":true,"conflictsOfInterestStatement":false,"humanSubjectEthicalGuidelines":false,"humanSubjectConsent":false,"humanSubjectClinicalTrial":false,"humanSubjectCaseReport":false,"vertebrateSubjectEthicalGuidelines":true},"doi":"10.21203/rs.3.rs-7832874/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7832874/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":93497298,"identity":"79c68791-3f4f-4e09-ac93-11e22528dd4a","added_by":"auto","created_at":"2025-10-14 13:21:59","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":6671314,"visible":true,"origin":"","legend":"","description":"","filename":"Funosasetal.worldassessmentBirdNET.docx","url":"https://assets-eu.researchsquare.com/files/rs-7832874/v1/ee710ae06a4cf7dbed9a6365.docx"},{"id":93496921,"identity":"e5e23bfd-75ec-4cb0-bfcc-622e0ae5abb3","added_by":"auto","created_at":"2025-10-14 13:13:59","extension":"json","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":342,"visible":true,"origin":"","legend":"","description":"","filename":"rs7832874.json","url":"https://assets-eu.researchsquare.com/files/rs-7832874/v1/b43be36c4c9b02db61f760fa.json"},{"id":93496924,"identity":"9429d41d-2212-4529-9c6b-029533ee26ca","added_by":"auto","created_at":"2025-10-14 13:13:59","extension":"xml","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1112555,"visible":true,"origin":"","legend":"","description":"","filename":"rs78328740enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-7832874/v1/a6dfda2723e078e68b8b21fe.xml"},{"id":93497300,"identity":"5990c590-4499-4a15-b18c-551429935559","added_by":"auto","created_at":"2025-10-14 13:21:59","extension":"png","order_by":3,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1109370,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-7832874/v1/3e0e0cd144f0d00593bb577b.png"},{"id":93496925,"identity":"4382159c-6c39-4998-ae79-ad983ac04815","added_by":"auto","created_at":"2025-10-14 13:13:59","extension":"jpeg","order_by":4,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":777864,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage2.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7832874/v1/f353cb998506325069251872.jpeg"},{"id":93498353,"identity":"aec0ec1b-6612-4817-9343-022a8e346a40","added_by":"auto","created_at":"2025-10-14 13:29:59","extension":"jpeg","order_by":5,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":169094,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage3.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7832874/v1/f5bebf3c2abc5a6ec0495b00.jpeg"},{"id":93496929,"identity":"0bb51753-d59f-430d-99f1-99073e2b5d7e","added_by":"auto","created_at":"2025-10-14 13:13:59","extension":"jpeg","order_by":6,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":2422989,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage4.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7832874/v1/ba0efbd1e2d988310774b5f2.jpeg"},{"id":93496927,"identity":"7614f834-c896-475c-a58f-53a727538f37","added_by":"auto","created_at":"2025-10-14 13:13:59","extension":"jpeg","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":341206,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage5.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7832874/v1/8f9184d9e46ec7de6bc16d25.jpeg"},{"id":93497299,"identity":"849d5e4c-bb45-4bdf-a896-cb0234d00e63","added_by":"auto","created_at":"2025-10-14 13:21:59","extension":"jpeg","order_by":8,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":200068,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage6.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7832874/v1/10774a1e53f8a6eacc2a4adb.jpeg"},{"id":93496939,"identity":"021b4703-6d09-413a-8e32-df585ebba7e9","added_by":"auto","created_at":"2025-10-14 13:13:59","extension":"jpeg","order_by":9,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":513513,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage7.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7832874/v1/3decf626c2090b58824bcf06.jpeg"},{"id":93497303,"identity":"ec0cab82-ad43-45dd-84fe-adb75f2040b7","added_by":"auto","created_at":"2025-10-14 13:21:59","extension":"png","order_by":10,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":211184,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-7832874/v1/3fb763201cbc1b3703e904e3.png"},{"id":93497301,"identity":"32eebd2a-cd6d-4841-9e9b-5fa11817e417","added_by":"auto","created_at":"2025-10-14 13:21:59","extension":"png","order_by":11,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":137924,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-7832874/v1/7b19918e03358a243b70c78d.png"},{"id":93496931,"identity":"52e7f846-e684-4d24-91fc-e3244932b771","added_by":"auto","created_at":"2025-10-14 13:13:59","extension":"png","order_by":12,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":74066,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-7832874/v1/6fee5e1e3a772bacd6fd047d.png"},{"id":93496933,"identity":"a9e48a27-4b32-46d3-b5fc-857cd9f435ac","added_by":"auto","created_at":"2025-10-14 13:13:59","extension":"png","order_by":13,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":421368,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-7832874/v1/a96a9f4aa72b6a96cca53bb1.png"},{"id":93496934,"identity":"144ee8b0-19ae-4221-8c27-5a939d919d6c","added_by":"auto","created_at":"2025-10-14 13:13:59","extension":"png","order_by":14,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":114326,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-7832874/v1/39317e9b1a626ba1c236ef2e.png"},{"id":93496938,"identity":"cab570af-4106-473b-b5ef-0984f7c594c1","added_by":"auto","created_at":"2025-10-14 13:13:59","extension":"png","order_by":15,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":103919,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-7832874/v1/3c4ef14e84f9ee15cc956ce0.png"},{"id":93497304,"identity":"135886fa-af29-444f-865a-3afa0a251aa8","added_by":"auto","created_at":"2025-10-14 13:21:59","extension":"png","order_by":16,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":181815,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage7.png","url":"https://assets-eu.researchsquare.com/files/rs-7832874/v1/5cc6c4490f6fc61b8c304904.png"},{"id":93496937,"identity":"fb657c8b-91ca-4a65-a28f-656f6b14c00d","added_by":"auto","created_at":"2025-10-14 13:13:59","extension":"xml","order_by":17,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1110103,"visible":true,"origin":"","legend":"","description":"","filename":"rs78328740structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-7832874/v1/c060b4d9b5c2f3d553e7e437.xml"},{"id":93496936,"identity":"2f1f107d-5440-4ae2-a5f7-1d947cbee7a0","added_by":"auto","created_at":"2025-10-14 13:13:59","extension":"html","order_by":18,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1090003,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-7832874/v1/49922c7a9682af392e8d9c0d.html"},{"id":93496918,"identity":"b817c870-f2c4-49b8-94b0-5d4218bb9488","added_by":"auto","created_at":"2025-10-14 13:13:58","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":1109370,"visible":true,"origin":"","legend":"\u003cp\u003e(A) Number of annotations (dot size), number of annotated species (radial distance from the center, with scales varying among continents), and proportion of annotated species not covered by BirdNET-Analyzer v2.4 (color scale) in each administrative region. Administrative regions (mostly countries) are grouped by continent*, each represented within a separate circle or distinguished by a different background color. (B) Global mapping of the proportion of species 1) correctly detected by BirdNET (green), 2) not detected by BirdNET (orange), and 3) not covered by BirdNET (purple), using a minimum confidence threshold of 0.1. * Hawai‘i and New Caledonia were both classified in Oceania following biogeographical criteria.\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-7832874/v1/d3487b8480a759b2e76a4638.png"},{"id":93496920,"identity":"e33a3e24-8564-4436-8986-5f52ebbff4d7","added_by":"auto","created_at":"2025-10-14 13:13:58","extension":"jpeg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":777864,"visible":true,"origin":"","legend":"\u003cp\u003eMean BirdNET precision and recall by administrative region using a minimum confidence threshold of 0.1. Point colors correspond to the continent of the administrative region. Results are shown separately at the (A) vocalisation and (B) dataset levels.\u003c/p\u003e","description":"","filename":"floatimage2.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7832874/v1/82358a7c6e9de3a7428ed640.jpeg"},{"id":93496917,"identity":"f5e7a105-665a-4042-bc1f-1522f8ec16fc","added_by":"auto","created_at":"2025-10-14 13:13:58","extension":"jpeg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":169094,"visible":true,"origin":"","legend":"\u003cp\u003eBirdNET Precision-Recall (PR) curves for each continent and level of analysis. Continents are ordered by decreasing vocalisation-level adj_AUC. Original Area Under the Curve (AUC) scores (orig_AUC) and AUC scores adjusted to account for recall range (adj_AUC) are shown on top of each curve. Curve shading reflects the confidence threshold: darker blue corresponds to lower thresholds and lighter blue to higher thresholds.\u003c/p\u003e","description":"","filename":"floatimage3.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7832874/v1/1c1a0f7aec1ee65219295887.jpeg"},{"id":93496923,"identity":"2cb4288d-a211-4bdd-bf99-3d3cc442f95e","added_by":"auto","created_at":"2025-10-14 13:13:59","extension":"jpeg","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":2422989,"visible":true,"origin":"","legend":"\u003cp\u003eProportion of true positives (TP, blue), false positives (FP, red), and false negatives (FN, yellow) by bird species in BirdNET output with minimum confidence scores of (A) 0.1, (B) 0.5, and (C) 0.75. Results are calculated at the vocalisation level, so the combined total of TPs and FNs per species is fixed (equal to the total number of vocalisations annotated), while FPs vary with confidence threshold, with lower thresholds yielding more predictions and thus more FPs. Only species present in at least 50 recordings are shown, listed in alphabetical order.\u003c/p\u003e","description":"","filename":"floatimage4.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7832874/v1/7a0874f359cb244ee4d2bb14.jpeg"},{"id":93498843,"identity":"c6578078-94b1-4272-a44d-a91c855df0a2","added_by":"auto","created_at":"2025-10-14 13:38:01","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":5922711,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7832874/v1/348e9154-bcc5-4f8c-92db-c8b64c8850f3.pdf"},{"id":93497297,"identity":"1d5c92f8-f945-47ac-bfbe-a85fc0505c16","added_by":"auto","created_at":"2025-10-14 13:21:58","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":2319500,"visible":true,"origin":"","legend":"","description":"","filename":"Supplementarymaterial.docx","url":"https://assets-eu.researchsquare.com/files/rs-7832874/v1/504a330f83c1ba4deb6a00f7.docx"}],"financialInterests":"The authors declare no competing interests.","formattedTitle":"\u003cp\u003e\u003cstrong\u003eA global assessment of BirdNET performance: differences among continents, biomes, and species\u003c/strong\u003e\u003c/p\u003e","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eIn recent decades, various automated and non-invasive approaches have become standard in biodiversity monitoring (Lahoz-Monfort and Magrath \u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). Among these, passive acoustic monitoring (PAM) has proven particularly effective for surveying diverse taxa, including anurans, bats, birds, cetaceans, and soniferous insects (Sugai et al. \u003cspan citationid=\"CR44\" class=\"CitationRef\"\u003e2019\u003c/span\u003e, Hoefer et al. \u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e2023\u003c/span\u003e, Darras et al. \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). Among terrestrial taxa, birds have been primary targets of PAM (Shonfield and Bayne \u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e2017\u003c/span\u003e, Sugai et al. \u003cspan citationid=\"CR44\" class=\"CitationRef\"\u003e2019\u003c/span\u003e), and soundscape ecology has historically drawn on acoustic ornithology (Gasc et al. \u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e2017\u003c/span\u003e). The documented capacity of this approach to characterise avian communities and estimate population densities from passively collected soundscapes (Darras et al. 2018, P\u0026eacute;rez-Granados and Traba \u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e2021\u003c/span\u003e) has prompted the development of dedicated methods and analytical pipelines for automated or semiautomated bird monitoring. In recent years, the development of effective low-cost audio recorders, such as the AudioMoth (Hill et al. \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e2018\u003c/span\u003e) and Song Meter Micro (Wildlife Acoustics) devices, along with essential advances in automated signal recognition software, have significantly expanded the use of PAM in biodiversity research.\u003c/p\u003e\u003cp\u003eWhile PAM offers numerous possibilities, it also presents challenges. One of its primary strengths\u0026mdash;the ability to easily scale acoustic monitoring both spatially and temporally\u0026mdash;also results in acoustic datasets far too large for manual analysis. To address this, most current projects rely on deep learning (DL) algorithms for the automated analysis of passively collected data (Stowell \u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e2022\u003c/span\u003e, Xie et al. 2022). Unfortunately, many state-of-the-art DL algorithms remain largely inaccessible to ecologists, land managers, and non-specialists lacking computational training. To bridge this gap, a new generation of user-friendly, ready-to-use sound recognition tools has emerged, helping to further streamline PAM workflows. Notably, many of these applications focus on birds as a model group, including BirdNET (Kahl et al. \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e2021\u003c/span\u003e), Perch (Ghani et al. \u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e2023\u003c/span\u003e), HawkEars (Huss et al. 2025), Nighthawk (Van Doren et al. \u003cspan citationid=\"CR48\" class=\"CitationRef\"\u003e2024\u003c/span\u003e), Chirpity (Kirkland \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2024\u003c/span\u003e), and the British Trust for Ornithology\u0026rsquo;s Acoustic Pipeline (BTO 2023). Among them, BirdNET stands out for its broad taxonomic and geographic coverage, as well as its high predictive accuracy (Kahl et al. \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e2021\u003c/span\u003e, P\u0026eacute;rez-Granados \u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e2023\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eBirdNET is a free bird vocalisation recognition software based on a convolutional neural network (Kahl et al. \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). Its latest version (v.2.4) can identify more than 6,000 bird species worldwide, as well as a smaller set of mammals and anurans (Wood et al. \u003cspan citationid=\"CR53\" class=\"CitationRef\"\u003e2023a\u003c/span\u003e, \u003cspan citationid=\"CR54\" class=\"CitationRef\"\u003e2023b\u003c/span\u003e, P\u0026eacute;rez-Granados et al. \u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e2023\u003c/span\u003e, Bota et al., \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). The model analyses recordings in 3-second windows, predicting zero, one, or multiple species per segment. Each prediction is assigned a confidence score ranging from 0.01 (very low model certainty in the prediction) to 1 (very high certainty), enabling users to filter results by using a customisable minimum confidence threshold. Low confidence thresholds favor high detection rates but increase the risk of false positives, while high confidence thresholds reduce errors at the cost of missed detections (Wood and Kahl \u003cspan citationid=\"CR52\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). Two additional parameters also shape BirdNET performance: \u003cem\u003eOverlap\u003c/em\u003e (ranging from 0 to 3 seconds), which determines the degree of temporal overlap between consecutive 3-second windows, and \u003cem\u003eSensitivity\u003c/em\u003e (ranging from 0.5 to 1.5), which adjusts how confidence scores are distributed across predictions. Lower \u003cem\u003eSensitivity\u003c/em\u003e values increase model certainty in its top predictions, while higher values make confidence scores more uniform across predictions (P\u0026eacute;rez-Granados et al. \u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e2025a\u003c/span\u003e). Recent studies have explored how these parameters influence BirdNET performance, with the optimal configuration differing among species, regions, and research goals (Funosas et al. \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2024\u003c/span\u003e, P\u0026eacute;rez-Granados et al. \u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e2025a\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eIn recent years, multiple studies have used BirdNET to automatically classify bird vocalisations and derive ecological or conservation insights from the data collected (e.g. Funosas et al., \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2024\u003c/span\u003e, wa Maina and Njoroge \u003cspan citationid=\"CR50\" class=\"CitationRef\"\u003e2025\u003c/span\u003e, Winiarska et al. \u003cspan citationid=\"CR51\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). Recent research suggests that BirdNET can provide a reliable characterisation of bird communities in Europe, often yielding higher species richness estimates than traditional on-site surveys due to its ability to detect nocturnal and cryptic species (Funosas et al. \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2024\u003c/span\u003e, Winiarska et al. \u003cspan citationid=\"CR51\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). However, most prior studies evaluating BirdNET have been limited in scope, focusing either on individual species at local scales (Manzano-Rubio et al. \u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e2022\u003c/span\u003e) or on datasets restricted to Europe and North America (P\u0026eacute;rez-Granados \u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). Only a handful of studies have evaluated BirdNET performance in other regions (Amor\u0026oacute;s-Ausina et al. \u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e2024\u003c/span\u003e, P\u0026eacute;rez-Granados and Schuchmann 2025). This geographic bias stems in part from the initial restriction of the BirdNET training set to European and North American species (Kahl et al. \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). Given the recent expansion of BirdNET species coverage and its increasing adoption worldwide, there is a growing need to rigorously evaluate its performance across diverse ecological, acoustic, and biogeographic contexts.\u003c/p\u003e\u003cp\u003eThe main goal of this study is to provide the first global evaluation of BirdNET performance. Using 4,224 one-minute soundscapes recorded at 67 sites across 28 administrative regions worldwide, we aim to assess how accurately BirdNET identifies bird vocalisations and how effectively it characterises bird communities from passively collected recordings. Our study has three objectives: (1) to examine geographic and biome-level variation in BirdNET capacity to identify individual vocalisations and characterise bird communities, (2) to provide species-level performance metrics by reporting mean precision (the proportion of correct predictions) and recall (the proportion of vocalisations successfully detected) values for each of the 1,102 bird species included in the study, and (3) to analyse how BirdNET performance varies across confidence thresholds. Given the rapidly growing use of BirdNET for automated bird monitoring, we hope that our findings will help identify its strengths and limitations across diverse ecological contexts, thus guiding future use of the software and contributing to its continued improvement.\u003c/p\u003e"},{"header":"2. Methods","content":"\u003cp\u003e\u003cstrong\u003e2.1. Acoustic data\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe recordings used in this study were sourced from the World Annotated Bird Acoustic Dataset (WABAD, version 2.0, P\u0026eacute;rez-Granados et al. 2025b). WABAD comprises expert-annotated recordings collected from 72 recording sites across 28 administrative regions (mainly representing countries). To facilitate further comparisons across datasets and ensure methodological consistency, we focused the study on a subset of 4,224 one-minute soundscapes from 67 distinct recording sites (henceforth referred to as datasets) in all 28 administrative regions, having retained only those with \u0026ldquo;strong labels\u0026rdquo; (i.e. providing the precise start and end times of each bird vocalisation). The spatial distribution of the acoustic datasets used is shown in Figure 1, and metadata for each site (minutes annotated, recording device, sampling frequency, geographic region, biome, and coordinates) are provided in Supplementary Table S1. Further details about the datasets analysed in the study are available in P\u0026eacute;rez-Granados et al. (2025c).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eData coverage in this study varied substantially across geographic regions, generally aligning with continental boundaries (Figure 1). The only exception is the Americas, which we divided into North America (Canada, USA, and Mexico) and Central/South America. This division allows us to evaluate whether BirdNET performance in North America\u0026mdash;where public training data are more abundant\u0026mdash;differs from that in the rest of the continent. For simplicity, we refer to geographic regions as \u0026ldquo;continents\u0026rdquo; hereafter. In Europe and Central/South America we gathered data from \u0026ge;20 datasets comprising \u0026gt;1,100 recordings and \u0026gt;27,000 annotated vocalisations each. In contrast, the other continents have far fewer annotated datasets (range 3\u0026ndash;9), recordings (range 181\u0026ndash;498), and annotated vocalisations (\u0026gt;6,000\u0026ndash;8,000; Figure 1). BirdNET species coverage, estimated as the proportion of bird species annotated by humans that was included in the algorithm, was near-complete in Europe and the Americas (\u0026ge;99%). Still, substantial gaps remained in Asia (95%), Africa (84%), and especially Oceania (72%). Table 1 summarises, for each continent, 1) the number of species and vocalisations annotated, and 2) the percentage of annotated species included in the BirdNET algorithm.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe ecological distribution of the recordings annotated was also uneven across biomes, categorised as in Olson et al. (2001). Tropical/subtropical and temperate broadleaf forests, as well as Mediterranean forests \u0026amp; shrublands, have 9,000\u0026ndash;35,000 vocalisations annotated. Boreal and temperate coniferous forests, tropical/subtropical grasslands, and wetlands are moderately represented (\u0026gt;3,000\u0026ndash;6,000 vocalisations annotated), while deserts \u0026amp; xeric shrublands, as well as temperate and montane grasslands \u0026amp; savannahs, have \u0026gt;500\u0026ndash;2,000 annotated vocalisations each. BirdNET species coverage was near-complete (\u0026ge;98%) in all biomes except for tropical/subtropical broadleaf forests (91%). Table 2 summarises, for each biome, 1) the number of annotated species and vocalisations, and 2) the percentage of annotated species included in the BirdNET algorithm.\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e2.2. Annotation procedure\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eExpert ornithologists with in-depth knowledge of the local bird communities annotated all recordings. Audio spectrograms were analysed in Raven Pro (v1.6), with experts being allowed to adjust the software parameters at their convenience and listen and visualise the recordings as many times as needed. Each vocalisation was labeled at the species level following the Clements Checklist nomenclature (Clements et al. 2021), the taxonomy used in BirdNET-Analyzer v.2.4 (Kahl et al. 2021), thus ensuring direct comparability between human annotations and BirdNET output. Each recording site was annotated by a single observer, who delineated vocalisations using bounding boxes encompassing their temporal and frequency ranges and exported the annotations as .txt files named after the corresponding recordings. Multiple vocalisations of the same species were grouped into a single annotation box if they occurred within one second of each other; otherwise, separate annotations were created. A more thorough description of the annotation workflow is available in P\u0026eacute;rez-Granados et al. (2025b), and both the recordings and annotations used in this study can be downloaded from P\u0026eacute;rez-Granados et al. (2025c).\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e2.3. Audio analysis\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWe processed the acoustic recordings using BirdNET-Analyzer v2.2.0 (model v2.4: BirdNET_GLOBAL_6K_V2.4_Model_FP32.tflite) via a Linux shell script interfaced with the algorithm\u0026rsquo;s Python backend, following the approach described in Funosas et al. (2024). BirdNET includes four adjustable settings that influence detection performance: \u003cem\u003eMinimum occurrence frequency threshold\u003c/em\u003e, \u003cem\u003eConfidence threshold\u003c/em\u003e, \u003cem\u003eOverlap\u003c/em\u003e, and \u003cem\u003eSensitivity\u003c/em\u003e (for detailed descriptions of these settings and their impact, see Kahl et al. 2021, Funosas et al. 2024, Wood and Kahl 2024, or P\u0026eacute;rez-Granados et al. 2025c). Based on prior research assessing nine combinations of \u003cem\u003eOverlap\u003c/em\u003e and \u003cem\u003eSensitivity\u003c/em\u003e values (P\u0026eacute;rez-Granados et al. 2025c), we selected an \u003cem\u003eOverlap\u003c/em\u003e of 2 seconds, maximising performance at both vocalisation and dataset levels, and a \u003cem\u003eSensitivity\u003c/em\u003e of 1 (default value). Because the effects of \u003cem\u003eSensitivity\u003c/em\u003e on BirdNET performance vary with continent and analytical goal (e.g. identifying individual vocalisations versus characterising bird communities), we opted for the default, balanced setting.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe \u003cem\u003eMinimum occurrence frequency threshold\u003c/em\u003e defines the lowest regional and temporal occurrence frequency a species must have to be included in the list of potentially detectable species generated by BirdNET (range: 0.01\u0026ndash;0.99). BirdNET-Analyzer v2.4 uses eBird checklist frequency data to estimate species ranges and probabilities of occurrence given geographic coordinates and week of the year (see https://github.com/birdnet-team/BirdNET-Analyzer/discussions/234). A low threshold broadens the list of potentially detectable species by including those with low likelihoods of occurrence, whereas a higher threshold restricts the list to species with the highest expected occurrence based on eBird data. We used a \u003cem\u003eMinimum occurrence frequency threshold\u003c/em\u003e of 0.02\u0026mdash;i.e. requiring a 2% probability of species presence for inclusion in BirdNET-generated species lists\u0026mdash;following Funosas et al. (2024), and retained the default \u003cem\u003eConfidence threshold\u003c/em\u003e of 0.1 to minimise the risk of missed detections and to evaluate how detection performance varies across different thresholds.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e2.4. BirdNET performance assessment\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWe evaluated BirdNET performance by comparing its predictions to expert annotations using a suite of custom R scripts (v4.2.2; R Core Team 2025) adapted, and available, from Funosas et al. (2024). Performance was assessed at two hierarchical levels: 1) the vocalisation level, providing detailed insight into BirdNET capacity to correctly identify individual songs or calls, and 2) the dataset level, reflecting its capacity to characterise bird community composition based on multiple recordings from a single recording site. BirdNET predictions were classified into four categories:\u003c/p\u003e\n\u003cul\u003e\n \u003cli\u003e\u003cstrong\u003eTrue Positives (TP):\u0026nbsp;\u003c/strong\u003eAt the vocalisation level, a BirdNET prediction was classified as a TP when an expert labelled the same species at the same 3-second time interval. At the dataset level, a bird species was considered a TP if there was at least one correct identification of that species by BirdNET in any of the recordings from the same study site (i.e. dataset).\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eFalse Positives (FP):\u003c/strong\u003e At the vocalisation level, a BirdNET prediction was classified as a FP when an expert did not detect the same species at the same time. At the dataset level, a bird species was considered a FP when all BirdNET predictions of that species in the dataset were incorrect.\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eTrue Negatives (TN):\u003c/strong\u003e A species was classified as a TN when it was covered by BirdNET but was neither identified by the expert nor predicted by the algorithm.\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eFalse Negatives (FN):\u003c/strong\u003e A species was classified as a FN when it was identified by the expert but not predicted by BirdNET.\u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003eBased on these categorisations, we evaluated BirdNET precision, recall and False Positive Rate (FPR) at both levels of analysis. Precision quantifies the proportion of correct identifications among all BirdNET predictions, whereas recall measures the proportion of expert-identified vocalisations or species that were correctly detected by BirdNET (P\u0026eacute;rez-Granados 2023). FPR complements these metrics by estimating the probability that BirdNET falsely detects a species absent from the acoustic sample\u0026mdash;defined as a 3-second prediction window at the vocalisation level and as the entire dataset at the dataset level\u0026mdash;and is computed as the number of spurious detections divided by the number of species absent from the sample but covered by BirdNET. Importantly, recall calculations include species not covered by BirdNET: any annotated species that went undetected was counted as a FN, whether because BirdNET failed to predict it despite coverage or because the species was outside the taxonomic scope of the algorithm. This choice was deliberate, as our concern is not BirdNET recall conditional on its species list, but the practical outcome for ecological applications\u0026mdash;i.e. if BirdNET is deployed in a region, what fraction of vocalisations or species it will detect successfully.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003ePrecision, recall and FPR were calculated across 90 confidence thresholds ranging from 0.10 to 0.99 in 0.01 increments, following Funosas et al. (2024). At the vocalisation level, recall was computed by aggregating all BirdNET predictions that overlapped a given vocalisation (i.e. those starting before its end and ending after its onset). Precision was calculated by pooling all expert annotations that overlapped each prediction. At the dataset level, comparisons were based on species lists: a species was marked as correctly predicted if it appeared in both BirdNET predictions and expert annotations, as long as they both coincided in time at some point. Note that under this criterion, a single correct detection of a species suffices for it to be classified as a TP at the dataset level, thus favouring higher recall values in longer datasets. The formulas used were the following:\u003c/p\u003e\n\u003cp\u003ePrecision\u0026nbsp;=\u0026nbsp;TP / (TP\u0026nbsp;+\u0026nbsp;FP)\u003c/p\u003e\n\u003cp\u003eRecall\u0026nbsp;=\u0026nbsp;TP / (TP\u0026nbsp;+\u0026nbsp;FN)\u003c/p\u003e\n\u003cp\u003eFPR = FP / (FP + TN)\u003c/p\u003e\n\u003cp\u003eTo visually represent the variation of these metrics across confidence thresholds, we used the Precision-Recall (PR) curve, accompanied by its corresponding Area Under the Curve (AUC; Davis and Goadrich, 2006). PR curves plot precision against recall across all confidence thresholds, capturing the trade-off between these two metrics, with higher AUC values (range 0\u0026ndash;1) being indicative of higher predictive power. Because PR AUC integrates precision across the entire recall range, broader recall ranges\u0026mdash;even those including lower recall values\u0026mdash;can yield higher AUCs. Thus, to enable fair comparisons across continents with varying recall ranges, we adjusted PR AUC scores to account for recall ranges using the following formula:\u003c/p\u003e\n\u003cp\u003e\u003cimg src=\"https://myfiles.space/user_files/58895_8739fc6c57c1c19a/58895_custom_files/img1760447169.png\" width=\"513\" height=\"79\"\u003e\u003c/p\u003e\n\u003cp\u003eFinally, we computed the F-score, which integrates precision and recall into a single performance metric:\u003c/p\u003e\n\u003cp\u003e\u003cimg src=\"https://myfiles.space/user_files/58895_8739fc6c57c1c19a/58895_custom_files/img1760447179.png\" width=\"666\" height=\"49\"\u003e\u003c/p\u003e\n\u003cp\u003eAn F-score with \u0026beta; = 1 gives equal weight to precision and recall, whereas \u0026beta; \u0026gt; 1 emphasises recall and \u0026beta; \u0026lt; 1 emphasises precision. We computed F-scores using three \u0026beta; values: \u0026beta; = 1 as a standard value to enable comparison with previous studies, \u0026beta; = 0.25 to prioritise precision over recall, and \u0026beta; = 4 to prioritise recall over precision. A \u0026beta; value \u0026lt; 1 was chosen because high precision is typically paramount in biodiversity research, where failing to detect a present species is generally less problematic than falsely detecting an absent one (Tolkova et al. 2021). Furthermore, specific population models (e.g. occupancy models) explicitly account for imperfect detection (Brunk et al. 2023, Bielski et al. 2024), further reducing the relative cost of FNs compared with FPs. A \u0026beta; value \u0026gt; 1, on the other hand, might be particularly relevant to research teams targeting rare or cryptic species and able to manually validate large numbers of BirdNET predictions. Moreover, some classification\u0026ndash;occupancy models can explicitly incorporate classification errors, making FPs less problematic (Ogawa et al 2025).\u003c/p\u003e\n\u003cp\u003eThe metrics described above were used to capture complementary aspects of BirdNET performance across continents, biomes, and species. PR AUC scores provided a global measure of predictive power, with PR curves offering a fine-grained view of how precision and recall shift with confidence threshold choice in each continent. F-score curves were used to identify, for every continent, the threshold that maximizes the trade-off between precision and recall under three weighting schemes: equal weighting, marked emphasis on precision, and marked emphasis on recall. To enable consistent cross-continent and cross-biome comparisons for different purposes\u0026mdash;whether identifying individual vocalisations (vocalisation level) or characterising bird communities (dataset level)\u0026mdash;, we calculated recall and precision at both levels of analysis using BirdNET\u0026rsquo;s default confidence threshold of 0.1. Two additional thresholds (0.5 and 0.75) were also used to evaluate how optimal threshold choice varies across species. We quantified uncertainty in continent- and biome-specific performance metrics using bootstrapped 95% confidence intervals across datasets within each continent and biome.\u003c/p\u003e"},{"header":"3. RESULTS","content":"\u003cp\u003e\u003cstrong\u003e3.1. Performance across continents\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWe found BirdNET performance to be moderately heterogeneous across continents. When using the default confidence threshold of 0.1, precision was relatively consistent across continents at the vocalisation level (range 0.57\u0026ndash;0.71, sd = 0.053), and more variable at the dataset level (range 0.27\u0026ndash;0.62, sd = 0.134; Figure 2), with Africa performing worst at both levels of analysis and Oceania performing best at the dataset level and second-to-best at the vocalisation level (Table 1). Recall, in turn, varied more widely across continents at the vocalisation level (range 0.24\u0026ndash;0.52, sd = 0.125) but was more uniform at the dataset level (range 0.49\u0026ndash;0.70, sd = 0.084), with Europe and North America performing consistently better and Africa and Asia performing consistently worse at both levels (Table 1, Figure 2, Supplementary Figure S1). Finally, FPR varied widely across continents at the vocalisation level (range 3e-05\u0026ndash;6e-05, sd = 1.03e-05) and even more so at the dataset level (range 0.001\u0026ndash;0.011, sd = 0.0034; Table 1), with high discrepancies in continent ranking between the two levels of analysis. Central and South America performed best and Europe performed worst at the vocalisation level, while Oceania performed best and Africa performed worst at the dataset level.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eAt the vocalisation level, adjusted PR AUC values highlight more evident differences: Asia (0.03) and Africa (0.04) exhibit very low scores at the vocalisation level, with Central/South America (0.13) and Europe (0.16) having intermediate scores, and Oceania (0.20) and North America (0.23) performing best. Asia shows the lowest adjusted PR AUC score because, even though the precision scores obtained with high confidence thresholds are very high, these drop precipitously when the threshold is lowered (Figure 3E). At the same time, the maximum recall score for this continent is already low (0.26), implying that any confidence threshold yielding reasonable precision will necessarily drive recall to very low levels. In Africa, both maximum precision and recall are lower than in Asia, but its precision demonstrates greater robustness, decreasing more slowly with reduced confidence thresholds (Figure 3A). At the dataset level, adjusted PR AUC values follow similar trends, but with a lower degree of cross-continent heterogeneity. Africa exhibits the lowest score (0.16), while Europe, Asia, Central/South America, and North America have intermediate scores (range 0.22\u0026ndash;0.29) and Oceania stands out (0.35) due to consistently high precision (\u0026ge;0.62) across confidence thresholds (Figure 3).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eF-score curves show broadly similar shapes across continents. F1-scores peak at a confidence threshold of 0.1 at the vocalisation level and plateau at a mostly flat maximum between 0.2 and 0.6 at the dataset level. F0.25-scores, for the most part, plateau at a relatively flat maximum between 0.3 and 0.8 at the vocalisation level and reach a peak at a confidence threshold of around 0.9 at the dataset level. F4-scores steadily decrease along with higher confidence thresholds at both levels of analysis, with declines being particularly pronounced at the dataset level (Figure S2).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e3.2. Performance across biomes\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eConsistent with continent-scale results, we found that, when using the default confidence threshold of 0.1, precision was relatively uniform across biomes at the vocalisation level (range 0.55\u0026ndash;0.76, sd = 0.065) but more heterogeneous at the dataset level (range 0.14\u0026ndash;0.41, sd = 0.090; Table 2). Best- and worst-performing biomes are not entirely consistent across the two levels: BirdNET performed especially poorly in wetlands at the vocalisation level and in deserts \u0026amp; xeric shrublands at the dataset level. In contrast, classifications in recordings from tropical/subtropical dry broadleaf forests and montane grasslands \u0026amp; savannahs exhibited the highest precision, maintaining strong performance at both levels of analysis.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eRecall varied more widely across biomes, with values of 0.39\u0026ndash;0.72 (sd = 0.105) at the vocalisation level and 0.51\u0026ndash;0.85 (sd = 0.127) at the dataset level (Table 2). Scores were consistently highest in deserts \u0026amp; xeric shrublands, whereas tropical/subtropical broadleaf forests ranked lowest at the vocalisation level and near the bottom at the dataset level, only surpassing montane grasslands \u0026amp; savannahs (Table 2, Figure S3). Finally, FPR varied widely across biomes at the vocalisation level (range 3e-05\u0026ndash;9e-05, sd = 1.91e-05) and even more so at the dataset level (range 0.003\u0026ndash;0.016, sd = 0.0038), with broad similarities in biome ranking between the two levels of analysis (Table 2). Montane grasslands \u0026amp; savannahs performed best at both levels, with temperate broadleaf \u0026amp; mixed forest and tropical/subtropical grasslands performing worst at the vocalisation and dataset levels, respectively. Confidence intervals for mean performance across both continents and biomes overlapped partially across all three metrics and both levels of analysis (Table 1), indicating moderate heterogeneity rather than clear-cut regional contrasts. Some of this overlap likely reflects uneven sample sizes across continents and biomes, as regions with fewer datasets exhibited broader confidence intervals and hence greater uncertainty in estimated means.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e3.3. Performance across species\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eOur species-specific analyses, focused on the species annotated in more than 50 recordings and conducted across three confidence thresholds (0.1, 0.5, 0.75), show that, as the minimum confidence threshold increases, both TPs and FPs decline, with FNs following the opposite trend (Figure 4). Consequently, the numbers of correctly and mistakenly detected species vary substantially with threshold choice. Because higher-confidence predictions are more likely to be correct, raising the threshold reduces FPs far more than TPs. This asymmetry, however, is highly species-dependent (Figure 4, Supplementary Table S2). Some species (e.g. \u003cem\u003eArremon brunneinucha\u003c/em\u003e) retain most TPs while eliminating nearly all FPs when increasing the confidence threshold from 0.1 to 0.75, achieving consistently high F1-scores (\u0026gt;0.7) across thresholds. However, other species (e.g. \u003cem\u003eAcrocephalus arundinaceus\u003c/em\u003e) lose \u0026ge;90% of TPs with the same confidence threshold increase, maintaining persistently low F1-scores (\u0026lt;0.3, Figure 4). Vocalisation-level precision and recall scores, along with the numbers of species-specific annotated vocalisations and BirdNET predictions for all 1,102 species in the WABAD dataset, are reported in Supplementary Table S2.\u003c/p\u003e"},{"header":"4. DISCUSSION","content":"\u003cp\u003eThe use of deep learning algorithms for automated wildlife detection from passively collected data has expanded rapidly in recent years (Stowell \u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e2022\u003c/span\u003e, Xie et al. 2022). Here, we provide the first comprehensive evaluation of BirdNET (Kahl et al. \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e2021\u003c/span\u003e) performance across continents and biomes, assessing both its ability to detect bird vocalisations and to characterise bird communities. We analysed an extensive acoustic dataset composed of 4,224 one-minute soundscapes annotated by local experts from 67 recording sites worldwide (P\u0026eacute;rez-Granados et al. \u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e2025a\u003c/span\u003e). Our results suggest that BirdNET cross-continent and cross-biome performances are highly heterogenous, with recall being more variable than precision. Much of this variation can be attributed to gaps in BirdNET species coverage, highlighting the importance of consulting the most up-to-date BirdNET species list when interpreting results. The fact that precision remains imperfect even at the highest confidence thresholds, and that relaxing these thresholds is unavoidable to obtain reasonable recall, makes expert validation indispensable for achieving accurate and comprehensive ecological inference.\u003c/p\u003e\u003cp\u003ePrecision exhibited substantial variation across continents and biomes, with vocalisation-level estimates being more consistent than those at the dataset level. Dataset-specific factors, including recording duration\u0026mdash;which tends to increase the number of FPs at fixed thresholds (Funosas et al. \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2024\u003c/span\u003e)\u0026mdash;and species richness\u0026mdash;where lower values magnify the relative impact of FPs\u0026mdash;likely contributed to this discrepancy. At the biome scale, recordings collected in wetland habitats exhibited the weakest performance at the vocalisation level. Since many wetland species are widespread and well represented in reference libraries, low performance possibly reflects ecological complexity driven by high species richness and abundance (Go\u0026euml;au et al. \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e2018\u003c/span\u003e), as well as the presence of species with relatively uncharacteristic or poorly differentiated calls (e.g. many waterfowl). Unlike passerines, where strong sexual selection and territoriality have driven the evolution of distinctive songs, waterfowl typically depend more on visual courtship displays, reducing pressure for acoustically distinctive signals (Johnsgard \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e1971\u003c/span\u003e, Ten Cate \u003cspan citationid=\"CR45\" class=\"CitationRef\"\u003e2021\u003c/span\u003e) and thereby complicating discrimination by BirdNET. Deserts, by contrast, exhibited the lowest dataset-level precision despite strong recall: with few species annotated per site, even a small number of incorrect detections can drastically reduce precision. At the continental scale, precision was found to be the weakest in Africa and Asia at both levels of analysis, possibly due to the relatively small amount of training data available for species occurring in these continents (Funosas et al. \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). Oceania, however, achieved the highest dataset-level precision and the second-highest vocalisation-level precision despite having the lowest degree of species coverage. This result may be partly explained by the partnership established between the Cornell Lab of Ornithology, developer of BirdNET, and the Listening Observatory for Hawaiian Ecosystems, possibly leading to the algorithm being better adjusted to the Hawaiian avifauna.\u003c/p\u003e\u003cp\u003eFPR broadly tracked precision, with vocalisation-level estimates more consistent than dataset-level ones. At the dataset level, continental rankings were nearly identical: Africa and Asia performed worst, Oceania and North America best. At the vocalisation level, however, Central and South America showed the lowest FPR, while Europe had the highest\u0026mdash;indicating that European precision is penalised mainly by its high number of FPs, whereas Central and South American precision is mainly driven down by its low number of TPs. At the biome scale, the biomes with the highest precision (tropical/subtropical dry broadleaf forests and montane grasslands \u0026amp; savannahs) also had the lowest FPR, while those with the highest FPR (temperate broadleaf \u0026amp; mixed forests and temperate coniferous forests at the vocalisation level; tropical/subtropical and temperate grasslands at the dataset level) also exhibited the lowest precision. This tight correspondence highlights the central role of FPs in shaping both metrics.\u003c/p\u003e\u003cp\u003eRecall also displayed broadly similar patterns to precision: continents and biomes with the most significant proportion of uncovered species (e.g. Africa, Asia, tropical/subtropical broadleaf forests) generally exhibited the lowest scores, with Oceania again standing out as an exception. Oceanian recall was surpassed only by Europe and North America and clearly outperformed Africa and Asia at both levels of analysis (Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). Beyond the potential emphasis on Hawaiian avifauna during BirdNET training, the low species richness in the geographically isolated Hawaiian bird communities\u0026mdash;accounting for half of our Oceanian data\u0026mdash;may also have contributed to higher detection rates by increasing the likelihood of detecting a large share of the present species. A similar pattern was observed in cross-biome analyses: deserts \u0026amp; xeric shrublands, which have the lowest species richness, achieved the highest recall at both levels of analysis (Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e). Taken together, these results underscore the combined influence of species coverage, dataset composition, and local context on BirdNET performance. However, given the limited number of datasets and uneven spatial coverage for certain continents and biomes\u0026mdash;most notably Oceania and deserts\u0026mdash;, these apparent patterns may partly reflect sampling artifacts rather than generalisable performance differences.\u003c/p\u003e\u003cp\u003eWhen integrating across the F1-score and PR AUC metrics, BirdNET performed best in North America, Europe, and Oceania, moderately in Central/South America, and relatively poorly in Asia and Africa, with these cross-continent patterns largely consistent between both levels of analysis. Geographic differences likely reflect disparities in training data: online acoustic libraries are richer in recordings from North America and Europe, whereas African and Asian species remain underrepresented (Macaulay \u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e2025\u003c/span\u003e, Xeno-canto \u003cspan citationid=\"CR56\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). The paucity of species-specific training data in these regions may predispose the algorithm to learn narrow or context-specific acoustic cues that fail to generalise more broadly. This could undermine the capacity of BirdNET to recognise species whose vocal signatures vary geographically or ecologically (Knight et al. \u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e2024\u003c/span\u003e, Sebasti\u0026aacute;n-Gonz\u0026aacute;lez and P\u0026eacute;rez-Granados \u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e2025\u003c/span\u003e), as well as species with very wide acoustic repertoires.\u003c/p\u003e\u003cp\u003eConfidence threshold analyses, in contrast, highlight a high degree of consistency across continents but a large difference between scales of evaluation. At the vocalisation level, F1-scores peak at a confidence threshold of 0.1 and decline steadily at higher thresholds, as recall decreases more sharply than precision improves. F0.25-scores plateau at intermediate thresholds because the greater weighting of precision over recall means that small precision gains can compensate for moderate recall losses. F4-scores, driven mainly by recall and only weakly by precision, decline steadily with higher confidence thresholds (Figure S2). At the dataset level, since only one correct detection is required for a species to be considered a TP, higher thresholds can improve results by favoring precision without as steep a recall penalty. This may explain why both recall rates and optimum thresholds are higher at the dataset level. These differences highlight that there is no universally optimal confidence threshold: the best choice is inherently context-dependent (Wood and Kahl \u003cspan citationid=\"CR52\" class=\"CitationRef\"\u003e2024\u003c/span\u003e, Tueng et al. 2025). Lower thresholds enhance recall and are therefore better suited to the monitoring of rare or elusive species, whereas higher thresholds are preferable for biodiversity monitoring, where minimizing FPs is essential to prevent inflated richness estimates. The instability of precision-focused F-scores at high thresholds also cautions against overconfidence, highlighting the need for manual validation\u0026mdash;especially in regions with low training coverage. More broadly, threshold selection is not just a technical choice but a reflection of ecological or conservation priorities. Explicitly reporting and justifying threshold choices could therefore enhance reproducibility and interpretability in future BirdNET applications.\u003c/p\u003e\u003cp\u003eOur species-level analyses also revealed substantial heterogeneity in BirdNET vocalisation-level performance across species (Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e4\u003c/span\u003e, Supplementary Table S2), a result that aligns with prior research suggesting strong variation across species, including within the same family (Amor\u0026oacute;s et al. 2024). Strong cross-context intraspecific variation in BirdNET performance has also been reported, with mean precision for the Common Raven (\u003cem\u003eCorvus corax\u003c/em\u003e) ranging from 0.29 (Cole et al. \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e2022\u003c/span\u003e) to 0.66 (Kahl \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2020\u003c/span\u003e) and 0.94 (Sethi et al. \u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). In our study, interspecific variability in vocalisation-level precision and recall diminished with increasing confidence thresholds, with standard deviations shrinking from 0.23 and 0.19 at a threshold of 0.1 to 0.18 and 0.13 at a threshold of 0.75. As expected, increasing thresholds improved precision but reduced recall, though the magnitude of this trade-off varied widely across species. In some cases, TPs declined far more slowly than FPs, making these species better suited to higher thresholds. In others, TPs and FPs declined at similar rates (Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e4\u003c/span\u003e), favoring lower thresholds. While the use of uniform thresholds for all species remains a practical option, these results underscore that species-specific adjustments are likely to yield more reliable outcomes (Wood and Kahl \u003cspan citationid=\"CR52\" class=\"CitationRef\"\u003e2024\u003c/span\u003e, Tseng et al. \u003cspan citationid=\"CR47\" class=\"CitationRef\"\u003e2025\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eDespite having limited our species-level analysis to those present in at least 50 recordings, there was substantial heterogeneity in the number of annotated vocalisations per species. Some species (e.g. \u003cem\u003eFringilla coelebs\u003c/em\u003e, \u003cem\u003eTurdus merula\u003c/em\u003e, \u003cem\u003eErithacus rubecula\u003c/em\u003e, \u003cem\u003eSylvia atricapilla\u003c/em\u003e) had more than 1000 annotated vocalisations across over 300 recordings, while others (e.g. \u003cem\u003eHerpsilochmus rufimarginatus\u003c/em\u003e, \u003cem\u003eCurruca undata\u003c/em\u003e, \u003cem\u003ePhasianus colchicus\u003c/em\u003e, \u003cem\u003eEurillas curvirostris\u003c/em\u003e) had fewer than 200 annotations drawn from only 50 recordings. Results for species with sparse representation\u0026mdash;whether in annotated vocalisations or BirdNET predictions\u0026mdash;should therefore be interpreted with particular caution. Moreover, several species, despite being frequent in the dataset, were largely undetected by BirdNET because they were excluded from many auto-generated species lists. A notable case is \u003cem\u003eAbroscopus albogularis\u003c/em\u003e, present in 969 recordings but included in only 111 lists and predicted in none. This result underlines the importance of revising the species lists automatically generated by BirdNET for each location and week of the year to detect potentially missing species. Since BirdNET generates these lists based on eBird data, this consideration will be especially important in undersampled regions, where lists are more likely to erroneously exclude locally present species (Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e, Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e). Further, prior work suggests that the performance of DL acoustic classifiers is not only species- but also context-dependent, influenced by factors such as background noise and the presence of acoustically similar taxa (Ventura et al. \u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e2024\u003c/span\u003e, Tseng et al. \u003cspan citationid=\"CR47\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). This underscores the need for prudence when extrapolating our species-specific results to datasets collected under different recording settings, background noise profiles, or bird community compositions (Wood and Kahl \u003cspan citationid=\"CR52\" class=\"CitationRef\"\u003e2024\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eInterpretation of our cross-continent and cross-biome results should also be undertaken cautiously, as the acoustic datasets analysed exhibit uneven representation across continents and biomes. Europe and Central/South America account for more than half of the data, whereas Africa, Asia, and Oceania are sparsely represented. Within continents, datasets are often concentrated in a few countries (e.g. Brazil and Colombia in Central/South America and Spain in Europe), so our findings cannot be assumed to be fully representative of entire world regions. This bias is particularly pronounced for Oceania, as our study covers only two regions in a continent with exceptional endemic diversity. Similarly, biome-level analyses reveal significant imbalances: tropical/subtropical, temperate, and Mediterranean forests are well represented, while deserts \u0026amp; xeric shrublands and montane grasslands \u0026amp; savannahs are poorly sampled in WABAD (P\u0026eacute;rez-Granados et al. 2025c).\u003c/p\u003e\u003cp\u003eA further limitation arises from the fact that continent and biome categories are not fully independent, since many biomes occur only in specific continents, complicating attribution. The partial overlap of cross-continent and cross-biome confidence intervals further suggests that the observed performance differences, while directionally consistent, may partly reflect these unequal sampling efforts rather than intrinsic model limitations. Consequently, apparent performance gaps should be interpreted with caution. In addition, the absence of common habitats such as croplands and urban areas, now central to biodiversity monitoring and research, limits the applicability of our findings to these environments. Additional sources of bias stem from the datasets themselves. Recording equipment and settings, annotation effort, and local conditions vary across locations, potentially influencing BirdNET performance (Leroy et al. 2021, Duc et al. \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e2021\u003c/span\u003e, Wood and Kahl \u003cspan citationid=\"CR52\" class=\"CitationRef\"\u003e2024\u003c/span\u003e, P\u0026eacute;rez-Granados \u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). Although all recordings were annotated by local experts following standardized protocols, differences in annotation quality remain possible, potentially biasing our results. Finally, we would like to acknowledge that although incorporating temporal and spatial filters (e.g., week and location) into the species list potentially detected by BirdNET is conceptually sound and improves ecological plausibility, it may inadvertently exclude species that are present but not expected according to reference databases, such as eBird. Such mismatches can affect detection performance and introduce seasonal or regional biases, affecting to a larger extent those areas with smaller reference databases. The impact of applying temporal and spatial filters on BirdNET performance should be further assessed to ensure robust and unbiased classification across contexts.\u003c/p\u003e\u003cp\u003eOverall, our study provides a practical framework for the informed use of BirdNET in biodiversity monitoring. While precision remains relatively consistent across regions, recall drops sharply in continents and biomes with limited acoustic data availability. Responsible BirdNET use requires objective- and, ideally, species-specific confidence thresholds, coupled with manual validation to ensure a rigorous and thorough characterisation of recorded bird communities. Our findings set the first global reference for these calibration decisions and emphasise the need to fill acoustic data gaps to ensure reliable, globally consistent model performance.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n \u003cli\u003eAmor\u0026oacute;s-Ausina, D., Schuchmann, K. L., Marques, M. I., \u0026amp; P\u0026eacute;rez-Granados, C. (2024). Living Together, Singing Together: Revealing Similar Patterns of Vocal Activity in Two Tropical Songbirds Applying BirdNET. \u003cem\u003eSensors\u003c/em\u003e, \u003cem\u003e24\u003c/em\u003e(17), 5780.\u003c/li\u003e\n \u003cli\u003eBielski, L., Cansler, C. A., McGinn, K., Peery, M. Z., \u0026amp; Wood, C. M. (2024). Can the Hermit Warbler (\u003cem\u003eSetophaga occidentalis\u003c/em\u003e) serve as an old-forest indicator species in the Sierra Nevada? \u003cem\u003eJournal of Field Ornithology, 95\u003c/em\u003e(1), 4.\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eBota, G., Manzano-Rubio, R., Catal\u0026aacute;n, L., G\u0026oacute;mez-Catas\u0026uacute;s, J., \u0026amp; P\u0026eacute;rez-Granados, C. (2023). Hearing to the unseen: AudioMoth and BirdNET as a cheap and easy method for monitoring cryptic bird species.\u0026nbsp;\u003cem\u003eSensors\u003c/em\u003e, \u003cem\u003e23\u003c/em\u003e(16), 7176.\u003c/li\u003e\n \u003cli\u003eBota, G., Manzano-Rubio, R., Fanlo, H., Franch, N., Brotons, L., Villero, D., ... \u0026amp; P\u0026eacute;rez-Granados, C. (2024). Passive acoustic monitoring and automated detection of the American bullfrog. \u003cem\u003eBiological Invasions\u003c/em\u003e, \u003cem\u003e26\u003c/em\u003e(4), 1269\u0026ndash;1279.\u003c/li\u003e\n \u003cli\u003eBiffi, S., Chapman, P. J., Engler, J. O., Kunin, W. E., \u0026amp; Ziv, G. (2024). Using automated passive acoustic monitoring to measure changes in bird and bat vocal activity around hedgerows of different ages. \u003cem\u003eBiological Conservation, 296,\u003c/em\u003e 110722.\u003c/li\u003e\n \u003cli\u003eBritish Trust for Ornithology. (2023). BTO Acoustic Pipeline: Detection and identification of birds, bats and wildlife from audio recordings. Available at: http://bto.org/pipeline. Last accessed 15/06/2025.\u003c/li\u003e\n \u003cli\u003eBrunk, K. M., Guti\u0026eacute;rrez, R. J., Peery, M. Z., Cansler, C. A., Kahl, S., \u0026amp; Wood, C. M. (2023). Quail on fire: Changing fire regimes may benefit mountain quail in fire-adapted forests. \u003cem\u003eFire Ecology, 19\u003c/em\u003e(1), 1\u0026ndash;13.\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eClements, J. F., Schulenberg, T. S., Iliff, M. J., Billerman, S. M., Fredericks, T. A., Gerbracht, J. A., Lepage, D., Sullivan, B. L., \u0026amp; Wood, C. L. (2021). The eBird/Clements checklist of Birds of the World: v2021.\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eCole, J. S., Michel, N. L., Emerson, S. A., \u0026amp; Siegel, R. B. (2022). Automated bird sound classifications of long-duration recordings produce occupancy model outputs similar to manually annotated data. \u003cem\u003eOrnithological Applications\u003c/em\u003e, \u003cem\u003e124\u003c/em\u003e(2), duac003.\u003c/li\u003e\n \u003cli\u003eCord, A. F., Darras, K., Ogawa, R., Barbaro, L., Gerling, C., Kernecker, M., Markova-Nenova, N., Rodriguez-Barrera, G., Zichner, F., \u0026amp; W\u0026auml;tzold, F. (2025). Leveraging passive acoustic monitoring for result-based agri-environmental schemes: Opportunities, challenges and next steps. \u003cem\u003eBiological Conservation, 305,\u003c/em\u003e 111042.\u003c/li\u003e\n \u003cli\u003eDarras, K., Bat\u0026aacute;ry, P., Furnas, B. J., Grass, I., Mulyani, Y. A., \u0026amp; Tscharntke, T. (2019). Autonomous sound recording outperforms human observation for sampling birds: a systematic map and user guide. \u003cem\u003eEcological Applications\u003c/em\u003e, \u003cem\u003e29\u003c/em\u003e(6), e01954.\u003c/li\u003e\n \u003cli\u003eDarras, K. F., Rountree, R., Van Wilgenburg, S., Cord, A. F., Chen, Y., Dong, L., ... \u0026amp; Wanger, T.C (2025). Worldwide Soundscapes: a synthesis of passive acoustic monitoring across realms. Global Ecology and Biogeography, 34(5): e70021.\u003c/li\u003e\n \u003cli\u003eDuc, P. N. H., Torterotot, M., Samaran, F., White, P. R., G\u0026eacute;rard, O., Adam, O., \u0026amp; Cazau, D. (2021). Assessing inter-annotator agreement from collaborative annotation campaign in marine bioacoustics. \u003cem\u003eEcological Informatics\u003c/em\u003e, 61, 101185.\u003c/li\u003e\n \u003cli\u003eFunosas, D., Barbaro, L., Schill\u0026eacute;, L., Elger, A., Castagneyrol, B., \u0026amp; Cauchoix, M. (2024). Assessing the potential of BirdNET to infer European bird communities from large-scale ecoacoustic data. \u003cem\u003eEcological Indicators\u003c/em\u003e, \u003cem\u003e164\u003c/em\u003e, 112146.\u003c/li\u003e\n \u003cli\u003eGasc, A., Francomano, D., Dunning, J. B., \u0026amp; Pijanowski, B. C. (2017). Future directions for soundscape ecology: The importance of ornithological contributions. \u003cem\u003eThe Auk\u003c/em\u003e, \u003cem\u003e134\u003c/em\u003e(1), 215\u0026ndash;228.\u003c/li\u003e\n \u003cli\u003eGhani, B., Denton, T., Kahl, S., \u0026amp; Klinck, H. (2023). Global birdsong embeddings enable superior transfer learning for bioacoustic classification. Scientific Reports, 13(1), 22876.\u003c/li\u003e\n \u003cli\u003eGo\u0026euml;au, H., Kahl, S., Glotin, H., Planqu\u0026eacute;, R., Vellinga, W.-P., \u0026amp; Joly, A. (2018). Overview of BirdCLEF 2018: Monospecies vs. soundscape bird identification. CEUR Workshops Proceedings, 2125(9).\u003c/li\u003e\n \u003cli\u003eHill, A. P., Prince, P., Pi\u0026ntilde;a Covarrubias, E., Doncaster, C. P., Snaddon, J. L., \u0026amp; Rogers, A. (2018). AudioMoth: Evaluation of a smart open acoustic device for monitoring biodiversity and the environment. \u003cem\u003eMethods in Ecology and Evolution\u003c/em\u003e, \u003cem\u003e9\u003c/em\u003e(5), 1199\u0026ndash;1211.\u003c/li\u003e\n \u003cli\u003eJohnsgard, P. A. (1971). Observations on sound production in the Anatidae. \u003cem\u003eWildfowl, 22\u003c/em\u003e, 46\u0026ndash;59.\u003c/li\u003e\n \u003cli\u003eKahl, S. (2020). Identifying Birds by Sound: Large-scale Acoustic Event Recognition for Avian Activity Monitoring. Unpublished doctoral dissertation, \u003cem\u003eChemnitz University of Technology\u003c/em\u003e.\u003c/li\u003e\n \u003cli\u003eKahl, S., Wood, C. M., Eibl, M., \u0026amp; Klinck, H. (2021). BirdNET: A deep learning solution for avian diversity monitoring. \u003cem\u003eEcological Informatics\u003c/em\u003e, \u003cem\u003e61\u003c/em\u003e, 101236.\u003c/li\u003e\n \u003cli\u003eKirkland, M. (2024). Chirpity‑Electron: AI‑powered audio analyser for bird call visualisation, detection and cataloguing Available at: https://github.com/Mattk70/Chirpity-Electron. Last accessed: 15/06/2025.\u003c/li\u003e\n \u003cli\u003eKnight, E., Rhinehart, T., de Zwaan, D. R., Weldy, M. J., Cartwright, M., Hawley, S. H., ... \u0026amp; Kitzes, J. (2024). Individual identification in acoustic recordings. \u003cem\u003eTrends in Ecology \u0026amp; Evolution\u003c/em\u003e, \u003cem\u003e39\u003c/em\u003e(10), 947\u0026ndash;960.\u003c/li\u003e\n \u003cli\u003eHoefer, S., McKnight, D. T., Allen-Ankins, S., Nordberg, E. J., \u0026amp; Schwarzkopf, L. (2023). Passive acoustic monitoring in terrestrial vertebrates: a review. \u003cem\u003eBioacoustics\u003c/em\u003e, \u003cem\u003e32\u003c/em\u003e(5), 506\u0026ndash;531.\u003c/li\u003e\n \u003cli\u003eLahoz-Monfort, J. J., \u0026amp; Magrath, M. J. (2021). A comprehensive overview of technologies for species and habitat monitoring and conservation. \u003cem\u003eBioScience\u003c/em\u003e, \u003cem\u003e71\u003c/em\u003e(10), 1038\u0026ndash;1062.\u003c/li\u003e\n \u003cli\u003eLassandro, M., Bolitho, L., \u0026amp; Newell, D. (2025). A systematic review of Passive Acoustic Monitoring for anuran conservation. \u003cem\u003eBioacoustics\u003c/em\u003e, 34(5), 579\u0026ndash;628.\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eLeroy, E. C., Thomisch, K., Royer, J. Y., Boebel, O., \u0026amp; Van Opzeeland, I. (2018). On the reliability of acoustic annotations and automatic detections of Antarctic blue whale calls under different acoustic conditions. \u003cem\u003eThe Journal of the Acoustical Society of America\u003c/em\u003e, \u003cem\u003e144\u003c/em\u003e(2), 740\u0026ndash;754.\u003c/li\u003e\n \u003cli\u003eMacaulay (2025). The World\u0026rsquo;s Premier Scientific Archive of Natural History Audio, Video, and Photographs. https://www.macaulaylibrary. org/about/.\u003c/li\u003e\n \u003cli\u003eManzano-Rubio, R., Bota, G., Brotons, L., Soto-Largo, E., \u0026amp; P\u0026eacute;rez-Granados, C. (2022). Low-cost open-source recorders and ready-to-use machine learning approaches provide effective monitoring of threatened species. \u003cem\u003eEcological Informatics\u003c/em\u003e, \u003cem\u003e72\u003c/em\u003e, 101910.\u003c/li\u003e\n \u003cli\u003eOgawa, R., Gosselin, F., Darras, K. F. A., Roilo, S., \u0026amp; Cord, A. F. (2025). A classification-occupancy model based on automatically identified species data. Ecology, 106(5), e70086.\u003c/li\u003e\n \u003cli\u003eOlson, D. M., Dinerstein, E., Wikramanayake, E. D., Burgess, N. D., Powell, G. V., Underwood, E. C., ... \u0026amp; Kassem, K. R. (2001). Terrestrial Ecoregions of the World: A New Map of Life on Earth: A new global map of terrestrial ecoregions provides an innovative tool for conserving biodiversity. BioScience, 51(11), 933-938.\u003c/li\u003e\n \u003cli\u003eP\u0026eacute;rez‐Granados, C. (2023). BirdNET: applications, performance, pitfalls and future opportunities. \u003cem\u003eIbis\u003c/em\u003e, \u003cem\u003e165\u003c/em\u003e(3), 1068-1075.\u003c/li\u003e\n \u003cli\u003eP\u0026eacute;rez-Granados, C. 2025. BirdNET\u0026rsquo;s confidence scores decrease with bird distance to the recorder: revisiting P\u0026eacute;rez-Granados (2023). Ardeola, \u003cem\u003e72\u003c/em\u003e(2): 149-159.\u003c/li\u003e\n \u003cli\u003eP\u0026eacute;rez‐Granados, C., \u0026amp; Traba, J. (2021). Estimating bird density using passive acoustic monitoring: a review of methods and suggestions for further research.\u0026nbsp;\u003cem\u003eIbis\u003c/em\u003e, \u003cem\u003e163\u003c/em\u003e(3), 765\u0026ndash;783.\u003c/li\u003e\n \u003cli\u003eP\u0026eacute;rez‐Granados, C., Bota, G., Giralt, D., Barrero, A., G\u0026oacute;mez‐Catas\u0026uacute;s, J., Bustillo‐De La Rosa, D., \u0026amp; Traba, J. (2019). Vocal activity rate index: a useful method to infer terrestrial bird abundance with acoustic monitoring.\u0026nbsp;\u003cem\u003eIbis\u003c/em\u003e, \u003cem\u003e161\u003c/em\u003e(4), 901\u0026ndash;907.\u003c/li\u003e\n \u003cli\u003eP\u0026eacute;rez-Granados, C., Feldman, M. J., \u0026amp; Mazerolle, M. J. (2023). Combining two user-friendly machine learning tools increases species detection from acoustic recordings.\u0026nbsp;\u003cem\u003eCanadian Journal of Zoology\u003c/em\u003e, \u003cem\u003e102\u003c/em\u003e(4), 403\u0026ndash;409.\u003c/li\u003e\n \u003cli\u003eP\u0026eacute;rez-Granados, C., Funosas, D., Morant, J., G\u0026oacute;mez, O.H.M., Mendoza, I., \u0026hellip; \u0026amp; Sebasti\u0026aacute;n-Gonz\u0026aacute;lez, E. 2025a. Optimisation of passive acoustic bird surveys: a global assessment of BirdNET settings. Preprint. Available at Research Square: DOI: 10.21203/rs.3.rs-6633549/v1.\u003c/li\u003e\n \u003cli\u003eP\u0026eacute;rez Granados, C., Morant, J., Darras, K.F.A., G\u0026oacute;mez, O.H.M., Mendoza, I., Mohedano-Mu\u0026ntilde;oz, M.A., \u0026hellip;. \u0026amp; Sebasti\u0026aacute;n-Gonz\u0026aacute;lez, E. 2025b. WABAD: A World Annotated Bird Acoustic Dataset for passive acoustic monitoring. Preprint. Available at. Research Square: DOI: 10.21203/rs.3.rs-5729784/v1\u003c/li\u003e\n \u003cli\u003eP\u0026eacute;rez Granados, C., Morant, J., Darras, K.F.A., G\u0026oacute;mez, O.H.M., Mendoza, I., Mohedano-Mu\u0026ntilde;oz, M.A., \u0026hellip;. \u0026amp; Sebasti\u0026aacute;n-Gonz\u0026aacute;lez, E. 2025c. WABAD: A World Annotated Bird Acoustic Dataset for passive acoustic monitoring. Zenodo. DOI: 10.5281/zenodo.15629388\u003c/li\u003e\n \u003cli\u003eSebasti\u0026aacute;n-Gonz\u0026aacute;lez, E., Camp, R. J., Tanimoto, A. M., de Oliveira, P. M., Lima, B. B., Marques, T. A., \u0026amp; Hart, P. J. (2018). Density estimation of sound-producing terrestrial animals using single automatic acoustic recorders and distance sampling. \u003cem\u003eAvian Conservation and Ecology 13\u003c/em\u003e(2), 7.\u003c/li\u003e\n \u003cli\u003eSethi, S. S., Foss\u0026oslash;y, F., Cretois, B., \u0026amp; Rosten, C. M. (2021). Management relevant applications of acoustic monitoring for Norwegian nature \u0026ndash; The Sound of Norway. NINA Report 2064. In 31. \u003cem\u003eNorsk institutt for naturforskning (NINA)\u003c/em\u003e.\u003c/li\u003e\n \u003cli\u003eShonfield, J., \u0026amp; Bayne, E. M. (2017). Autonomous recording units in avian ecological research: current use and future applications. \u003cem\u003eAvian Conservation \u0026amp; Ecology\u003c/em\u003e, \u003cem\u003e12\u003c/em\u003e(1), 14.\u003c/li\u003e\n \u003cli\u003eStowell, D. (2022). Computational bioacoustics with deep learning: a review and roadmap.\u0026nbsp;\u003cem\u003ePeerJ\u003c/em\u003e, \u003cem\u003e10\u003c/em\u003e, e13152.\u003c/li\u003e\n \u003cli\u003eSugai, L. S. M., Silva, T. S. F., Ribeiro Jr, J. W., \u0026amp; Llusia, D. (2019). Terrestrial passive acoustic monitoring: review and perspectives. \u003cem\u003eBioScience\u003c/em\u003e, \u003cem\u003e69\u003c/em\u003e(1), 15\u0026ndash;25.\u003c/li\u003e\n \u003cli\u003eten Cate, C. (2021). Re-evaluating vocal production learning in non-oscine birds. \u003cem\u003ePhilosophical Transactions of the Royal Society B: Biological Sciences\u003c/em\u003e, 376(1836), 20200249.\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eTolkova, I., Chu, B., Hedman, M., Kahl, S., \u0026amp; Klinck, H. (2021). Parsing Birdsong with Deep Audio Embeddings. arXiv.\u003c/li\u003e\n \u003cli\u003eTseng, S., Hodder, D. P., \u0026amp; Otter, K. A. (2025). Setting BirdNET confidence thresholds: species-specific vs. universal approaches. \u003cem\u003eJournal of Ornithology\u003c/em\u003e, 1\u0026ndash;13.\u003c/li\u003e\n \u003cli\u003eVan Doren, B. M., Farnsworth, A., Stone, K., Osterhaus, D. M., Drucker, J., \u0026amp; Van Horn, G. (2024). Nighthawk: acoustic monitoring of nocturnal bird migration in the Americas. \u003cem\u003eMethods in Ecology and Evolution\u003c/em\u003e, \u003cem\u003e15\u003c/em\u003e(2), 329\u0026ndash;344.\u003c/li\u003e\n \u003cli\u003eVentura, T. M., Ganchev, T. D., P\u0026eacute;rez-Granados, C., De Oliveira, A. G., de SG Pedroso, G., Marques, M. I., \u0026amp; Schuchmann, K. L. (2024). The importance of acoustic background modelling in CNN-based detection of the neotropical White-lored Spinetail (Aves, Passeriformes, Furnaridae). \u003cem\u003eBioacoustics, 33\u003c/em\u003e(2), 103\u0026ndash;121.\u003c/li\u003e\n \u003cli\u003ewa Maina, C., \u0026amp; Njoroge, P. (2025). Comparing point counts, passive acoustic monitoring, citizen science and machine learning for bird species monitoring in the Mount Kenya ecosystem. \u003cem\u003ePhilosophical Transactions B\u003c/em\u003e, \u003cem\u003e380\u003c/em\u003e(1928), 20240057.\u003c/li\u003e\n \u003cli\u003eWiniarska, D., Neubauer, G., Budka, M., Szymański, P., Barczyk, J., Cholewa, M., \u0026amp; Osiejuk, T. S. (2025). BirdNET provides superior diversity estimates compared to observer-based surveys in long-term monitoring. \u003cem\u003eEcological Indicators\u003c/em\u003e, \u003cem\u003e177\u003c/em\u003e, 113747.\u003c/li\u003e\n \u003cli\u003eWood, C. M., \u0026amp; Kahl, S. (2024). Guidelines for appropriate use of BirdNET scores and other detector outputs. \u003cem\u003eJournal of Ornithology\u003c/em\u003e, 165: 777\u0026ndash;782.\u003c/li\u003e\n \u003cli\u003eWood, C. M., Kahl, S., Barnes, S., Van Horne, R., \u0026amp; Brown, C. (2023a). Passive acoustic surveys and the BirdNET algorithm reveal detailed spatiotemporal variation in the vocal activity of two anurans. \u003cem\u003eBioacoustics\u003c/em\u003e, \u003cem\u003e32\u003c/em\u003e(5), 532\u0026ndash;543.\u003c/li\u003e\n \u003cli\u003eWood, C. M., Barceinas Cruz, A., \u0026amp; Kahl, S. (2023b). Pairing a user‐friendly machine‐learning animal sound detector with passive acoustic surveys for occupancy modeling of an endangered primate. \u003cem\u003eAmerican Journal of Primatology\u003c/em\u003e, \u003cem\u003e85\u003c/em\u003e(8), e23507.\u003c/li\u003e\n \u003cli\u003eXie, J., Zhong, Y., Zhang, J., Liu, S., Ding, C., \u0026amp; Triantafyllopoulos, A. (2023). A review of automatic recognition technology for bird vocalizations in the deep learning era. \u003cem\u003eEcological Informatics\u003c/em\u003e, \u003cem\u003e73\u003c/em\u003e, 101927.\u003c/li\u003e\n \u003cli\u003eXeno-canto (2025). Sharing Bird Sounds from Around the World. https://www.xeno-canto.org/about/xeno-canto.\u003c/li\u003e\n\u003c/ol\u003e"},{"header":"Tables","content":"\u003cp\u003eTable 1: BirdNET performance across continents when using a minimum confidence threshold of 0.1. For each continent, the table reports the number of annotated vocalisations and species, the number and proportion of annotated species 1) correctly detected, 2) in BirdNET auto-generated lists but not detected, 3) covered but not in BirdNET auto-generated lists, and 4) not covered by BirdNET, as well as precision, recall, and FPR (reported as mean \u0026plusmn; standard deviation across datasets, followed by the bootstrapped 95% confidence interval) calculated at both vocalisation (voc_precision, voc_recall, voc_FPR) and dataset (ds_precision, ds_recall, ds_FPR) levels.\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"691\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eContinent\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eVocalisations annotated\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eSpecies annotated\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eSpecies correctly detected\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eSpecies in auto-generated lists but not detected\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eSpecies covered but not included in auto-generated lists\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eSpecies not covered\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 55px;\"\u003e\n \u003cp\u003e\u003cstrong\u003evoc_precision\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eds_precision\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 52px;\"\u003e\n \u003cp\u003e\u003cstrong\u003evoc_recall\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 52px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eds_recall\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 55px;\"\u003e\n \u003cp\u003e\u003cstrong\u003evoc_FPR\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eds_FPR\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 53px;\"\u003e\n \u003cp\u003eAfrica\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 53px;\"\u003e\n \u003cp\u003e6455\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 53px;\"\u003e\n \u003cp\u003e169\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e82 (49%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e55 (33%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e5 (3%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e27 (16%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 55px;\"\u003e\n \u003cp\u003e0.568 \u0026plusmn; 0.156 (0.445\u0026ndash;0.715)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e0.272 \u0026plusmn; 0.095 (0.185\u0026ndash;0.348)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 52px;\"\u003e\n \u003cp\u003e0.236 \u0026plusmn; 0.084 (0.156\u0026ndash;0.299)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 52px;\"\u003e\n \u003cp\u003e0.487 \u0026plusmn; 0.079 (0.42\u0026ndash;0.546)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 55px;\"\u003e\n \u003cp\u003e4e-05 \u0026plusmn; 3e-05 (2.04e-05\u0026ndash;6.93e-05)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e0.011 \u0026plusmn; 0.004 (0.00706\u0026ndash;0.0133)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 53px;\"\u003e\n \u003cp\u003eAsia\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 53px;\"\u003e\n \u003cp\u003e8410\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 53px;\"\u003e\n \u003cp\u003e106\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e62 (58%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e28 (26%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e11 (10%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e5 (5%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 55px;\"\u003e\n \u003cp\u003e0.68 \u0026plusmn; 0.22 (0.444\u0026ndash;0.878)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e0.274 \u0026plusmn; 0.096 (0.163\u0026ndash;0.336)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 52px;\"\u003e\n \u003cp\u003e0.257 \u0026plusmn; 0.123 (0.183\u0026ndash;0.338)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 52px;\"\u003e\n \u003cp\u003e0.527 \u0026plusmn; 0.151 (0.37\u0026ndash;0.672)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 55px;\"\u003e\n \u003cp\u003e5e-05 \u0026plusmn; 2e-05 (1.84e-05\u0026ndash;6.38e-05)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e0.009 \u0026plusmn; 0.005 (0.00323\u0026ndash;0.0125)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 53px;\"\u003e\n \u003cp\u003eCentral and South America\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 53px;\"\u003e\n \u003cp\u003e27582\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 53px;\"\u003e\n \u003cp\u003e482\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e269 (56%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e178 (37%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e28 (6%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e7 (1%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 55px;\"\u003e\n \u003cp\u003e0.707 \u0026plusmn; 0.118 (0.654\u0026ndash;0.758)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e0.344 \u0026plusmn; 0.132 (0.288\u0026ndash;0.407)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 52px;\"\u003e\n \u003cp\u003e0.458 \u0026plusmn; 0.179 (0.383\u0026ndash;0.536)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 52px;\"\u003e\n \u003cp\u003e0.59 \u0026plusmn; 0.161 (0.524\u0026ndash;0.659)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 55px;\"\u003e\n \u003cp\u003e3e-05 \u0026plusmn; 2e-05 (2.55e-05\u0026ndash;4.06e-05)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e0.007 \u0026plusmn; 0.005 (0.00503\u0026ndash;0.00909)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 53px;\"\u003e\n \u003cp\u003eEurope\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 53px;\"\u003e\n \u003cp\u003e29781\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 53px;\"\u003e\n \u003cp\u003e182\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e149 (82%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e20 (11%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e11 (6%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e2 (1%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 55px;\"\u003e\n \u003cp\u003e0.605 \u0026plusmn; 0.167 (0.543\u0026ndash;0.669)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e0.282 \u0026plusmn; 0.1 (0.243\u0026ndash;0.32)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 52px;\"\u003e\n \u003cp\u003e0.512 \u0026plusmn; 0.165 (0.443\u0026ndash;0.572)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 52px;\"\u003e\n \u003cp\u003e0.692 \u0026plusmn; 0.159 (0.634\u0026ndash;0.751)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 55px;\"\u003e\n \u003cp\u003e6e-05 \u0026plusmn; 3e-05 (4.87e-05\u0026ndash;7.22e-05)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e0.008 \u0026plusmn; 0.003 (0.00653\u0026ndash;0.00861)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 53px;\"\u003e\n \u003cp\u003eNorth America\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 53px;\"\u003e\n \u003cp\u003e8748\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 53px;\"\u003e\n \u003cp\u003e164\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e124 (76%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e37 (23%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e2 (1%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e1 (1%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 55px;\"\u003e\n \u003cp\u003e0.647 \u0026plusmn; 0.164 (0.54\u0026ndash;0.745)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e0.371 \u0026plusmn; 0.124 (0.297\u0026ndash;0.446)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 52px;\"\u003e\n \u003cp\u003e0.517 \u0026plusmn; 0.255 (0.353\u0026ndash;0.657)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 52px;\"\u003e\n \u003cp\u003e0.696 \u0026plusmn; 0.174 (0.587\u0026ndash;0.809)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 55px;\"\u003e\n \u003cp\u003e5e-05 \u0026plusmn; 5e-05 (2.36e-05\u0026ndash;8.49e-05)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e0.005 \u0026plusmn; 0.003 (0.00352\u0026ndash;0.00649)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 53px;\"\u003e\n \u003cp\u003eOceania\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 53px;\"\u003e\n \u003cp\u003e8085\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 53px;\"\u003e\n \u003cp\u003e54\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e32 (59%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e5 (9%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e2 (4%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e15 (28%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 55px;\"\u003e\n \u003cp\u003e0.692 \u0026plusmn; 0.131 (0.572\u0026ndash;0.78)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e0.624 \u0026plusmn; 0.23 (0.426\u0026ndash;0.823)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 52px;\"\u003e\n \u003cp\u003e0.448 \u0026plusmn; 0.358 (0.166\u0026ndash;0.73)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 52px;\"\u003e\n \u003cp\u003e0.611 \u0026plusmn; 0.292 (0.341\u0026ndash;0.81)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 55px;\"\u003e\n \u003cp\u003e5e-05 \u0026plusmn; 5e-05 (1.02e-05\u0026ndash;9.17e-05)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e0.001 \u0026plusmn; 0.001 (0.000384\u0026ndash;0.00161)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTable 2: BirdNET performance across biomes when using a minimum confidence threshold of 0.1. For each biome, the table reports the number of annotated vocalisations and species, the number and proportion of annotated species 1) correctly detected, 2) in BirdNET auto-generated lists but not detected, 3) covered by BirdNET but not included in BirdNET auto-generated lists, and 3) not covered by BirdNET, as well as precision, recall, and FPR (reported as mean \u0026plusmn; standard deviation across datasets, followed by the bootstrapped 95% confidence interval) calculated at both vocalisation (voc_precision, voc_recall, voc_FPR) and dataset (ds_precision, ds_recall, ds_FPR) levels. Standard deviations and confidence intervals are not provided for biomes being represented by a single dataset.\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"689\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eBiome\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eVocalisations annotated\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eSpecies annotated\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eSpecies correctly detected\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eSpecies in auto-generated lists but not detected\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eSpecies covered but not included in auto-generated lists\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eSpecies not covered\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e\u003cstrong\u003evoc_precision\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eds_precision\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e\u003cstrong\u003evoc_recall\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eds_recall\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 55px;\"\u003e\n \u003cp\u003e\u003cstrong\u003evoc_FPR\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 51px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eds_FPR\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 53px;\"\u003e\n \u003cp\u003eBoreal Forest/Taiga\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 53px;\"\u003e\n \u003cp\u003e5604\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 53px;\"\u003e\n \u003cp\u003e79\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e56 (71%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e23 (29%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e0 (0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e0 (0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e0.586 \u0026plusmn; 0.191 (0.445\u0026ndash;0.748)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e0.347 \u0026plusmn; 0.088 (0.281\u0026ndash;0.42)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e0.411 \u0026plusmn; 0.149 (0.275\u0026ndash;0.521)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e0.636 \u0026plusmn; 0.117 (0.545\u0026ndash;0.73)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 55px;\"\u003e\n \u003cp\u003e5e-05 \u0026plusmn; 4e-05 (2.72e-05\u0026ndash;8.01e-05)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 51px;\"\u003e\n \u003cp\u003e0.005 \u0026plusmn; 0.003 (0.00268\u0026ndash;0.00767)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 53px;\"\u003e\n \u003cp\u003eDeserts \u0026amp; Xeric Shrublands\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 53px;\"\u003e\n \u003cp\u003e623\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 53px;\"\u003e\n \u003cp\u003e13\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e11 (85%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e2 (15%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e0 (0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e0 (0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e0.675\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e0.143\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e0.72\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e0.846\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 55px;\"\u003e\n \u003cp\u003e4e-05\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 51px;\"\u003e\n \u003cp\u003e0.01\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 53px;\"\u003e\n \u003cp\u003eMediterranean Forests \u0026amp; Shrublands\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 53px;\"\u003e\n \u003cp\u003e10572\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 53px;\"\u003e\n \u003cp\u003e116\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e79 (68%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e27 (23%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e8 (7%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e2 (2%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e0.641 \u0026plusmn; 0.138 (0.555\u0026ndash;0.717)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e0.218 \u0026plusmn; 0.104 (0.165\u0026ndash;0.29)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e0.528 \u0026plusmn; 0.153 (0.434\u0026ndash;0.611)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e0.633 \u0026plusmn; 0.165 (0.536\u0026ndash;0.729)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 55px;\"\u003e\n \u003cp\u003e5e-05 \u0026plusmn; 2e-05 (3.85e-05\u0026ndash;5.85e-05)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 51px;\"\u003e\n \u003cp\u003e0.009 \u0026plusmn; 0.003 (0.00739\u0026ndash;0.0103)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 53px;\"\u003e\n \u003cp\u003eMontane Grasslands \u0026amp; Savannahs\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 53px;\"\u003e\n \u003cp\u003e529\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 53px;\"\u003e\n \u003cp\u003e32\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e19 (59%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e12 (38%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e1 (3%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e0 (0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e0.761 \u0026plusmn; 0.014 (0.751\u0026ndash;0.771)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e0.357 \u0026plusmn; 0.017 (0.345\u0026ndash;0.368)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e0.452 \u0026plusmn; 0.086 (0.392\u0026ndash;0.513)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e0.507 \u0026plusmn; 0.044 (0.476\u0026ndash;0.538)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 55px;\"\u003e\n \u003cp\u003e3e-05 \u0026plusmn; 0 (3.13e-05\u0026ndash;3.65e-05)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 51px;\"\u003e\n \u003cp\u003e0.003 \u0026plusmn; 0.001 (0.00292\u0026ndash;0.00369)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 53px;\"\u003e\n \u003cp\u003eTemperate Broadleaf \u0026amp; Mixed Forest\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 53px;\"\u003e\n \u003cp\u003e10099\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 53px;\"\u003e\n \u003cp\u003e118\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e96 (81%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e21 (18%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e1 (1%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e0 (0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e0.598 \u0026plusmn; 0.166 (0.476\u0026ndash;0.705)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e0.373 \u0026plusmn; 0.064 (0.331\u0026ndash;0.419)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e0.547 \u0026plusmn; 0.269 (0.346\u0026ndash;0.712)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e0.855 \u0026plusmn; 0.099 (0.788\u0026ndash;0.928)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 55px;\"\u003e\n \u003cp\u003e9e-05 \u0026plusmn; 5e-05 (5.07e-05\u0026ndash;0.000125)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 51px;\"\u003e\n \u003cp\u003e0.006 \u0026plusmn; 0.002 (0.00539\u0026ndash;0.00758)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 53px;\"\u003e\n \u003cp\u003eTemperate Coniferous Forest\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 53px;\"\u003e\n \u003cp\u003e3611\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 53px;\"\u003e\n \u003cp\u003e50\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e43 (86%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e5 (10%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e2 (4%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e0 (0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e0.644 \u0026plusmn; 0.15 (0.539\u0026ndash;0.75)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e0.401 \u0026plusmn; 0.034 (0.377\u0026ndash;0.425)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e0.657 \u0026plusmn; 0.149 (0.552\u0026ndash;0.762)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e0.878 \u0026plusmn; 0.093 (0.812\u0026ndash;0.944)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 55px;\"\u003e\n \u003cp\u003e8e-05 \u0026plusmn; 6e-05 (4.18e-05\u0026ndash;0.00012)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 51px;\"\u003e\n \u003cp\u003e0.005 \u0026plusmn; 0.002 (0.00431\u0026ndash;0.00662)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 53px;\"\u003e\n \u003cp\u003eTemperate Grasslands\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 53px;\"\u003e\n \u003cp\u003e2107\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 53px;\"\u003e\n \u003cp\u003e34\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e24 (71%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e8 (24%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e2 (6%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e0 (0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e0.672\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e0.245\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e0.533\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e0.706\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 55px;\"\u003e\n \u003cp\u003e6e-05\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 51px;\"\u003e\n \u003cp\u003e0.011\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 53px;\"\u003e\n \u003cp\u003eTropical/Subtropical Grasslands\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 53px;\"\u003e\n \u003cp\u003e4281\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 53px;\"\u003e\n \u003cp\u003e91\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e57 (63%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e31 (34%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e1 (1%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e2 (2%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e0.676 \u0026plusmn; 0.056 (0.636\u0026ndash;0.716)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e0.221 \u0026plusmn; 0.005 (0.217\u0026ndash;0.224)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e0.441 \u0026plusmn; 0.172 (0.319\u0026ndash;0.562)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e0.626 \u0026plusmn; 0.099 (0.556\u0026ndash;0.696)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 55px;\"\u003e\n \u003cp\u003e6e-05 \u0026plusmn; 1e-05 (5.33e-05\u0026ndash;7.26e-05)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 51px;\"\u003e\n \u003cp\u003e0.016 \u0026plusmn; 0.002 (0.0139\u0026ndash;0.0171)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 53px;\"\u003e\n \u003cp\u003eTropical/Subtropical Dry Broadleaf Forest\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 53px;\"\u003e\n \u003cp\u003e9334\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 53px;\"\u003e\n \u003cp\u003e191\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e118 (62%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e50 (26%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e5 (3%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e18 (9%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e0.753 \u0026plusmn; 0.072 (0.704\u0026ndash;0.802)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e0.415 \u0026plusmn; 0.199 (0.291\u0026ndash;0.556)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e0.387 \u0026plusmn; 0.269 (0.212\u0026ndash;0.575)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e0.575 \u0026plusmn; 0.198 (0.44\u0026ndash;0.701)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 55px;\"\u003e\n \u003cp\u003e3e-05 \u0026plusmn; 2e-05 (1.52e-05\u0026ndash;4.03e-05)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 51px;\"\u003e\n \u003cp\u003e0.004 \u0026plusmn; 0.003 (0.00216\u0026ndash;0.00574)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 53px;\"\u003e\n \u003cp\u003eTropical/Subtropical Moist Broadleaf Forest\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 53px;\"\u003e\n \u003cp\u003e35687\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 53px;\"\u003e\n \u003cp\u003e605\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e331 (55%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e197 (33%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e42 (7%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e35 (6%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e0.659 \u0026plusmn; 0.155 (0.594\u0026ndash;0.724)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e0.37 \u0026plusmn; 0.165 (0.306\u0026ndash;0.439)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e0.409 \u0026plusmn; 0.208 (0.324\u0026ndash;0.496)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e0.577 \u0026plusmn; 0.169 (0.509\u0026ndash;0.641)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 55px;\"\u003e\n \u003cp\u003e4e-05 \u0026plusmn; 3e-05 (2.66e-05\u0026ndash;4.9e-05)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 51px;\"\u003e\n \u003cp\u003e0.006 \u0026plusmn; 0.004 (0.00484\u0026ndash;0.00813)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 53px;\"\u003e\n \u003cp\u003eWetland\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 53px;\"\u003e\n \u003cp\u003e6614\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 53px;\"\u003e\n \u003cp\u003e183\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e106 (58%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e72 (39%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e5 (3%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e0 (0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e0.547 \u0026plusmn; 0.226 (0.394\u0026ndash;0.711)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e0.271 \u0026plusmn; 0.068 (0.227\u0026ndash;0.318)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e0.505 \u0026plusmn; 0.14 (0.415\u0026ndash;0.604)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 53px;\"\u003e\n \u003cp\u003e0.62 \u0026plusmn; 0.127 (0.536\u0026ndash;0.7)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 55px;\"\u003e\n \u003cp\u003e6e-05 \u0026plusmn; 2e-05 (4.12e-05\u0026ndash;7.34e-05)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 51px;\"\u003e\n \u003cp\u003e0.009 \u0026plusmn; 0.005 (0.00556\u0026ndash;0.0117)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"Forest Science and Technology Center of Catalonia","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"passive acoustic monitoring, bird communities, BirdNET, deep learning, automated detection, confidence threshold","lastPublishedDoi":"10.21203/rs.3.rs-7832874/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7832874/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eRecent advances in machine learning have accelerated automated species detection across diverse ecological domains, enabling large-scale, non-invasive monitoring of biodiversity. In ornithological research, coupling passive acoustic monitoring (PAM) with rapidly-developing novel identification tools such as BirdNET\u0026mdash;a deep learning\u0026ndash;based sound recognition algorithm\u0026mdash;offers new opportunities for surveying vocally active bird communities. Yet, BirdNET performance across diverse ecological and biogeographic contexts remains to be quantified. Here, we present the first worldwide evaluation of BirdNET using 4,224 one-minute soundscapes from 67 sites across 28 administrative regions annotated by local experts that included 1,020 species. More specifically, we assessed the capacity of BirdNET to correctly identify individual vocalisations and characterise bird communities based on the automated analysis of passively collected soundscapes. We further analysed how its performance varies across continents, biomes, species, and minimum confidence thresholds. The proportion of correct BirdNET predictions (precision) was generally high and consistent across continents (range: 0.57\u0026ndash;0.71 at the vocalisation level) and biomes (range: 0.55\u0026ndash;0.76 at the vocalisation level). In contrast, the proportion of vocalisations or species successfully detected (recall) was generally lower and more heterogeneous across continents (range: 0.24\u0026ndash;0.52 at the vocalisation level) and biomes (range: 0.34\u0026ndash;0.72 at the vocalisation level), reflecting differences in species coverage and local ecological context. BirdNET predictive power, as measured by the Precision-Recall Area Under the Curve (PR AUC), was highest in North America, Oceania, and Europe (range: 0.16\u0026ndash;0.23 at the vocalisation level), moderate in Central/South America (0.13), and lowest in Africa and Asia (range: 0.03\u0026ndash;0.04). Species-specific analyses revealed substantial heterogeneity in detection accuracy, with optimal confidence thresholds varying widely by species and analytical goal. Our results establish a global reference point for BirdNET reliability and highlight where algorithmic refinement and expanded acoustic sampling are most needed.\u003c/p\u003e","manuscriptTitle":"A global assessment of BirdNET performance: differences among continents, biomes, and species","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-10-14 13:13:54","doi":"10.21203/rs.3.rs-7832874/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"9a3b8c89-d462-48d9-ac27-c1c434f7b146","owner":[],"postedDate":"October 14th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":56126066,"name":"Artificial Intelligence and Machine Learning"}],"tags":[],"updatedAt":"2025-10-14T13:13:54+00:00","versionOfRecord":[],"versionCreatedAt":"2025-10-14 13:13:54","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7832874","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7832874","identity":"rs-7832874","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}