Search

Funding: Wolf Creek Foundation; National Science Foundation. Grant Numbers: CCF-1522054, CNS-1059284, DBI-1939187, ICER-1927646; National Aeronautics and Space Administration. Grant Number: NNH12ZDA001N-ECOF; Leon Levy Foundation. Aim Ecological data collected by the general public are valuable for addressing a wide range of ecological research and conservation planning, and there has been a rapid increase in the scope and volume of data available. However, data from eBird or other large-scale projects with volunteer observers typically present several challenges that can impede robust ecological inferences. These challenges include spatial bias, variation in effort and species reporting bias. Innovation We use the example of estimating species distributions with data from eBird, a community science or citizen science (CS) project. We estimate two widely used metrics of species distributions: encounter rate and occupancy probability. For each metric, we critically assess the impact of data processing steps that either degrade or refine the data used in the analyses. CS data density varies widely across the globe, so we also test whether differences in model performance are robust to sample size. Main conclusions Model performance improved when data processing and analytical methods addressed the challenges arising from CS data; however, the degree of improvement varied with species and data density. The largest gains we observed in model performance were achieved with 1) the use of complete checklists (where observers report all the species they detect and identify, allowing non-detections to be inferred) and 2) the use of covariates describing variation in effort and detectability for each checklist. Occupancy models were more robust to a lack of complete checklists. Improvements in model performance with data refinement were more evident with larger sample sizes. In general, we found that the value of each refinement varied by situation and we encourage researchers to assess the benefits in other scenarios. These approaches will enable researchers to more effectively harness the vast ecological knowledge that exists within CS data for conservation and basic research. Postprint

The steady decline of avian populations worldwide urgently calls for a cyber-physical system to monitor bird migration at the continental scale. Compared to other sources of information (radar and crowdsourced observations), bioacoustic sensor networks combine low latency with a high taxonomic specificity. However, the scarcity of flight calls in bioacoustic monitoring scenes (below 0.1% of total recording time) requires the automation of audio content analysis. In this article, we address the problem of scaling up the detection and classification of flight calls to a full-season dataset: 6672 hours across nine sensors, yielding around 480 million neural network predictions. Our proposed pipeline, BirdVox, combines multiple machine learning modules to produce per-species flight call counts. We evaluate BirdVox on an annotated subset of the full season (296 hours) and discuss the main sources of estimation error which are inherent to a real-world deployment: mechanical sensor failures, sensitivity to background noise, misdetection, and taxonomic confusion. After developing dedicated solutions to mitigate these sources of error, we demonstrate the usability of BirdVox by reporting a species-specific temporal estimate of flight call activity for the Swainson’s Thrush (Catharus ustulatus).

Computer and information scientists join forces with other fields to help solve societal and environmental challenges facing humanity, in pursuit of a sustainable future.

Citizen science data are increasingly making valuable contributions to ecological studies. However, many citizen science surveys are also designed to encourage wide participation and therefore the participants have a range of natural history expertise, leading to variation and potentially bias in the data. We assessed a recently proposed measure of observer expertise, calculated based on the average numbers of species recorded by observers. We investigated if this observer expertise score is associated with how often an observer reports any individual species. Species reporting rates increased monotonically with the observer's expertise score for 197 of 200 species, suggesting that this expertise score describes inter-observer variation in the detectability of individual species. Expertise scores were incorporated into single-species occupancy models as a covariate, to explain inter-observer variation in detectability. Including expertise as a detectability covariate led to improved model fit and improved predictive performance on validation data. The expertise score had a large effect on the estimated detectability, comparable in magnitude to the effect of the duration of the observation period. Expertise scores were also included into single-species occupancy models that estimated seasonal patterns in species occupancy and seasonal expertise effects. The addition of a seasonal effect of expertise led to improved model fit and increased predictive performance on validation data. The seasonal expertise accounted for bias that may be introduced by seasonal differences in the effect of expertise, caused by changes in the environment or species behaviour. Measures of observer expertise included in models as a covariate can account for heterogeneity and bias introduced by variable expertise, although in this example the differences in estimated occupancy were small. This method of incorporating observer expertise can be used in any regression model of species occurrence, occupancy, abundance, or density to produce more reliable ecological inference and may be most important where citizen science schemes encourage wide participation. Overall, the results highlight the value of recording observer identity and other detectability covariates, to control for sources of bias associated with the observation process.

Ensuring that conservation decisions are informed by the best available data is a fundamental challenge in the face of rapid global environmental change. Too often, new science is not easily or quickly translated into conservation action. Traditional approaches to data collection and science delivery may be both inefficient and insufficient, as conservation practitioners need access to salient, credible, and legitimate data to take action. Open access data could serve as a tool to help bridge the gap between science and action, by providing conservation practitioners with access to relevant data in near real time. Broad-scale citizen-science data represent a fast-growing resource for open access databases, providing relevant and appropriately scaled data on organisms, much in the way autonomous sensors do so on the environment. Several such datasets are now broadly available, yet documentation of their application to conservation is rare. Here we use eBird, a project where individuals around the world submit data on bird distribution and abundance, as an example of how citizen-science data can be used to achieve tangible conservation science and action at local, regional, and global scales. Our examination illustrates how these data can be strategically applied to improve our understanding of spatial and temporal distributions of birds, the impacts of anthropogenic change on ecological systems, and creative conservation solutions to complex problems. We raise awareness of the types of conservation action now happening with citizen-science data, and discuss the benefits, limitations, and caveats of this approach.

Bioacoustic sensors, sometimes known as autonomous recording units (ARUs), can record sounds of wildlife over long periods of time in scalable and minimally invasive ways. Deriving per-species abundance estimates from these sensors requires detection, classification, and quantification of animal vocalizations as individual acoustic events. Yet, variability in ambient noise, both over time and across sensors, hinders the reliability of current automated systems for sound event detection (SED), such as convolutional neural networks (CNN) in the time-frequency domain. In this article, we develop, benchmark, and combine several machine listening techniques to improve the generalizability of SED models across heterogeneous acoustic environments. As a case study, we consider the problem of detecting avian flight calls from a ten-hour recording of nocturnal bird migration, recorded by a network of six ARUs in the presence of heterogeneous background noise. Starting from a CNN yielding state-of-the-art accuracy on this task, we introduce two noise adaptation techniques, respectively integrating short-term (60 milliseconds) and long-term (30 minutes) context. First, we apply per-channel energy normalization (PCEN) in the time-frequency domain, which applies short-term automatic gain control to every subband in the mel-frequency spectrogram. Secondly, we replace the last dense layer in the network by a context-adaptive neural network (CA-NN) layer. Combining them yields state-of-the-art results that are unmatched by artificial data augmentation alone. We release a pre-trained version of our best performing system under the name of BirdVoxDetect, a ready-to-use detector of avian flight calls in field recordings., 32 pages, in English. Submitted to PLOS ONE journal in February 2019; revised August 2019; published October 2019

Biodiversity is being lost at an unprecedented rate, and monitoring is crucial for understanding the causal drivers and assessing solutions. Most biodiversity monitoring data are collected by volunteers through citizen science projects, and often crucial information is lacking to account for the inevitable biases that observers introduce during data collection. We contend that citizen science projects intended to support biodiversity monitoring must gather information about the observation process as well as species occurrence. We illustrate this using eBird, a global citizen science project that collects information on bird occurrences as well as vital contextual information on the observation process while maintaining broad participation. Our fundamental argument is that regardless of what species are being monitored, when citizen science projects collect a small set of basic information about how participants make their observations, the scientific value of the data collected will be dramatically improved.

Citizen science data are valuable for addressing a wide range of ecological research questions, and there has been a rapid increase in the scope and volume of data available. However, data from large-scale citizen science projects typically present a number of challenges that can inhibit robust ecological inferences. These challenges include: species bias, spatial bias, and variation in effort.To demonstrate addressing key challenges in analysing citizen science data, we use the example of estimating species distributions with data from eBird, a large semi-structured citizen science project. We estimate two widely applied metrics of species distributions: encounter rate and occupancy probability. For each metric, we assess the impact of data processing steps that either degrade or refine the data used in the analyses. We also test whether differences in model performance are maintained at different sample sizes.Model performance improved when data processing and analytical methods addressed the challenges arising from citizen science data. The largest gains in model performance were achieved with: 1) the use of complete checklists (where observers report all the species they detect and identify); and 2) the use of covariates describing variation in effort and detectability for each checklist. Occupancy models were more robust to a lack of complete checklists and effort variables. Improvements in model performance with data refinement were more evident with larger sample sizes.Here, we describe processes to refine semi-structured citizen science data to estimate species distributions. We demonstrate the value of complete checklists, which can inform the design and adaptation of citizen science projects. We also demonstrate the value of information on effort. The methods we have outlined are also likely to improve other forms of inference, and will enable researchers to conduct robust analyses and harness the vast ecological knowledge that exists within citizen science data.

Most spatial conservation planning for wide-ranging or migratory species is constrained by poor knowledge of species' spatiotemporal dynamics and is only based on static species' ranges. However, species have substantial variation in abundance across their range and migratory species have important spatiotemporal population dynamics. With growing ecological data and advancing analytics, both of these can be estimated and incorporated into spatial conservation planning. However, there is limited information on the degree to which including this information affects conservation planning. We compared the performance of systematic conservation prioritizations for different scenarios based on varying the input species' distributions by ecological metric (abundance distributions versus range maps) and temporal sampling resolution (weekly, monthly, or quarterly). We used the example of a community of 41 species of migratory shorebirds that breed in North America, and we used eBird data to produce weekly estimates of species' abundances and ranges. Abundance distributions at a monthly or weekly resolution led to prioritizations that most efficiently protected species throughout the full annual cycle. Conversely, spatial prioritizations based on species' ranges required more sites and left most species insufficiently protected for at least part of their annual cycle. Prioritizations with only quarterly species ranges were very inefficient as they needed to target 40% of species' ranges to include 10% of populations. We highlight the high value of abundance information for spatial conservation planning, which leads to more efficient and effective spatial prioritization for conservation. Overall, we provide evidence that spatial conservation planning for wide-ranging migratory species is most robust and efficient when informed by species' abundance information from the full annual cycle.

While many observations of species are being collected by citizen science projects worldwide, it can be challenging to identify projects collecting data that effectively monitor biodiversity. Over the past several years the allure of taking a “Big Data” approach has provided the opportunity to gather massive quantities of observations via the Internet, too often with insufficient information to describe how the observations were made. Information about species populations — where and when they occur and how many of them are there — (i.e., the signal) can be lost because insufficient information is gathered to account for the inherent biases in data collection (i.e., the noise). Here we suggest that citizen science projects that have succeeded in motivating large numbers of participants, must consider factors that influence the ecological process that affect species populations as well as the observation process that determines how observations are made. Those citizen science projects that collect sufficient contextual information describing the observation process can be used to generate increasingly accurate information about the distribution and abundance of organisms. We illustrate this using eBird as a case study, describing how this citizen science platform is able to collect vital contextual information on the observation process while maintaining a broad global constituency of participants. We highlight how eBird provides information with which to generate biodiversity indicators — specifically distribution, abundance, and habitat associations — across the entire annual cycle, even for populations of long distance migratory birds, a highly challenging taxon.

eBird is a global citizen science project that gathers observations of birds. The project has been making a considerable contribution to the collection and sharing of bird observations, even in the data-poorest countries, and is accelerating the accumulation of bird records globally. On 22 March 2018 eBird surpassed ½ billion bird observations. A primary component of ensuring the best quality data is the network of more than 1300 volunteer reviewers who scour incoming data for accuracy. Reviewers provide active feedback to participants on everything from bird identification to best practices for data collection. Since eBird’s inception in 2002, almost 23 million observations have been reviewed, requiring more than 190,000 hours of effort by reviewers. In this presentation we review how eBird recruits expert reviewers, describe their responsibilities, and offer some insight in new developments to improve the reviewing process. How are reviewers recruited. There are three primary methods that used to identify new reviewers. First, if we don’t have any active participants in a region (e.g., Kamchatka Russia) eBird staff search birding listserves to find an individual who is reporting a lot of high-quality observations from the area. We then contact those individuals and offer them the opportunity to review records for the region. This option has the lowest likelihood of success. Second, if an individual is submitting a lot of records to eBird from a region that needs a reviewer we contact them and request their participation. Third, in much of the world eBird has partner groups. These partner organizations (e.g., Taiwan, Spain, India, Portugal, Australia, and all of the Western Hemisphere) recruit their own reviewers. The third method is the most effective way to gain expert participation. What does a reviewer do? eBird reviewers work to improve eBird data in three primary areas. First, they develop and manage the eBird checklist filters for a region. These filters generate a checklist of birds for a particular time and location, and determine what records get flagged for further review. Second, if an eBird participant tries to report a species that is not on the checklist, or if the number of individuals of a species exceeds the filter limit, then these records get flagged for review. Reviewers contact the observer and request further documentation. Currently, 57% of all records that are evaluated by reviewers are validated. Finally, eBird reviewers validate whether the participant is eBirding correctly. That is, are they correctly filling out the information on when, where, and how they went birding. It has been our experience that different types of reviewers are required to effectively review eBird submissions: those who are good at reviewing bird records and those who are good at educating observers on how to participate. What are future plans? eBird will move towards more effective reviewer teams, where the volume of observations can be split amongst a number of individuals with different strengths, allowing identification experts to focus on observation-level ID issues; and strong communicators to focus on working with contributors on checklist-level best practices. Currently, a single eBird review platform handles a broad array of different reviewing functions. It is our intent to split some of these functions into multiple platforms. For example, right now all review happens at the database level of the ‘observation’: a record of a taxon at a date and location. Plans are underway to develop tools that will allow reviewers to work at the entire checklist level (i.e., to more easily review the accuracy of how all the observations during a checklist event were submitted), which will enable much more effective review of checklist-level data quality concerns.

This article addresses the automatic detection of vocal, nocturnally migrating birds from a network of acoustic sensors. Thus far, owing to the lack of annotated continuous recordings, existing methods had been benchmarked in a binary classification setting (presence vs. absence). Instead, with the aim of comparing them in event detection, we release BirdVox-full-night, a dataset of 62 hours of audio comprising 35402 flight calls of nocturnally migrating birds, as recorded from 6 sensors. We find a large performance gap between energy-based detection functions and data-driven machine listening. The best model is a deep convolutional neural network trained with data augmentation. We correlate recall with the density of flight calls over time and frequency and identify the main causes of false alarm.

Avian migration is one of Earth's largest processes of biomass transport, involving billions of birds. We estimated continental biomass flows of nocturnal avian migrants across the contiguous United States using a network of 143 weather radars. We show that, relative to biomass leaving in autumn, proportionally more biomass returned in spring across the southern United States than across the northern United States. Neotropical migrants apparently achieved higher survival during the combined migration and non-breeding period, despite an average three- to fourfold longer migration distance, compared with a more northern assemblage of mostly temperate-wintering migrants. Additional mortality expected with longer migration distances was probably offset by high survival in the (sub)tropics. Nearctic-Neotropical migrants relying on a 'higher survivorship' life-history strategy may be particularly sensitive to variations in survival on the overwintering grounds, highlighting the need to identify and conserve important non-breeding habitats.

Over the next 5 years major advances in the development and application of numerous technologies related to computing, mobile phones, artificial intelligence (AI), and augmented reality (AR) will have a dramatic impact in biodiversity monitoring and conservation. Over a 2-week period several of us had the opportunity to meet with multiple technology experts in the Silicon Valley, California, USA to discuss trends in technology innovation, and how they could be applied to conservation science and ecology research. Here we briefly highlight some of the key points of these meetings with respect to AI and Deep Learning. Computing: Investment and rapid growth in AI and Deep Learning technologies are transforming how machines can perceive the environment. Much of this change is due to increased processing speeds of Graphics Processing Units (GPUs), which is now a billion-dollar industry. Machine learning applications, such as convolutional neural networks (CNNs) run more efficiently on GPUs and are being applied to analyze visual imagery and sounds in real time. Rapid advances in CNNs that use both supervised and unsupervised learning to train the models is improving accuracy. By taking a Deep Learning approach where the base layers of the model are built upon datasets of known images and sounds (supervised learning) and later layers relying on unclassified images or sounds (unsupervised learning), dramatically improve the flexibility of CNNs in perceiving novel stimuli. The potential to have autonomous sensors gathering biodiversity data in the same way personal weather stations gather atmospheric information is close at hand. Mobile Phones: The phone is the most widely used information appliance in the world. No device is on the near horizon to challenge this platform, for several key reasons. First, network access is ubiquitous in many parts of the world. Second, batteries are improving by about 20% annually, allowing for more functionality. Third, app development is a growing industry with significant investment in specializing apps for machine-learning. While GPUs are already running on phones for video streaming, there is much optimism that reduced or approximate Deep Learning models will operate on phones. These models are already working in the lab, with the biggest hurdle being power consumption and developing energy efficient applications and algorithms to run complicated AI processes will be important. It is just a matter of time before industry will have AI functionality on phones. These rapid improvements in computing and mobile phone technologies have huge implications for biodiversity monitoring, conservation science, and understanding ecological systems. Computing: AI processing of video imagery or acoustic streams create the potential to deploy autonomous sensors in the environment that will be able to detect and classify organisms to species. Further, AI processing of Earth spectral imagery has the potential to provide finer grade classification of habitats, which is essential in developing fine scale models of species distributions over broad spatial and temporal extents. Mobile Phones: increased computing functionality and more efficient batteries will allow applications to be developed that will improve an individual’s perception of the world. Already AI functionality of Merlin improves a birder’s ability to accurately identify a bird. Linking this functionality to sensor devices like specialized glasses, binoculars, or listening devises will help an individual detect and classify objects in the environment. In conclusion, computing technology is advancing at a rapid rate and soon autonomous sensors placed strategically in the environment will augment the species occurrence data gathered by humans. The mobile phone in everyone’s pocket should be thought of strategically, in how to connect people to the environment and improve their ability to gather meaningful biodiversity information.

Species-level observational data comprise the largest and fastest-growing part of the Global Biodiversity Information Facility (GBIF). The largest single contributor of species observations is eBird, which so far has contributed more than 361 million records to GBIF. eBird engages a vast network of human observers (citizen-scientists) to report bird observations, with the goal of estimating the range, abundance, habitat preferences, and trends of bird species at high spatial and temporal resolutions across each species’ entire life-cycle. Since its inception, eBird has focused on improving the data quality of its observations, primarily focused in two areas: ensuring that participants describe how they gathered their observations and, all observations are reviewed for accuracy. In this presentation I will review how this is done in eBird. Standardized Data Collection. eBird gathers bird observations based on how bird watchers typically observe birds with units of data collection being “checklists” of zero or more species including a count of individuals for each species observed. Participants choose the location where they made their observations and submit their checklists via Mobile Apps (50% of all submissions) or the website (50% of all submissions). All checklists are submitted in a standard format identifying where, how, and with whom they made their observations. Mobile apps precisely record locations, the track taken, and the distance they traveled while making the observations. The start time and duration of surveys are also recorded. All observers must report whether they reported all the birds they detected and identified, which allows analysts to infer absence of birds if they were not reported. All data are stored within an Oracle data management framework. Data Accuracy. The most significant data quality challenge for species observations is detecting and correctly identifying organisms to species. The issue involves how to handle both false positives — the misidentification of an observed organism, and false negatives—failing to report a species that was present. The most egregious false positives can be identified as anomalies that fall outside the norm of occurrence for a species at a particular time or space. However, false positives can also be misidentifications of common species. These challenges are addressed by: Data-driven filters. eBird’s existing data can identify and flag potentially erroneous records at increasingly fine spatial, temporal, and user-specific scales. These filters can identify outliers and likely errors, which are the foundation of the eBird review process. By using the vetted data to identify outliers, data quality checks run against expected occurrence probabilities at very fine scales and identify anomalies during data submission (including on mobile devices). Incorporate observer expertise scores. Observer differences are the largest source of variability in eBird data. Assessment of observer metrics, and the inclusion of these data in species distribution models, improves analysis output and model performance. Expert reviewer network. More than 2000 volunteers review records identified by the data-driven filters and contact data submitters to confirm their observations. The existing data quality process functions globally. Currently the approach is focused on misidentified birds, but in the future will also involve collection event issues (e.g., issues with protocol, location, or methodology), sensitive species, exotic species, and better handle widely-observed individual rarities. Additional tools are also to be developed to help editors improve efficiency and better prioritize review. In 2017, 4,107,757 observations representing 4.6% of all eBird records submitted were flagged for review by the data driven filters. Of these records 57.4% were validated and 42.6% were invalidated.

Much biodiversity data is collected worldwide, but it remains challenging to assemble the scattered knowledge for assessing biodiversity status and trends. The concept of Essential Biodiversity Variables (EBVs) was introduced to structure biodiversity monitoring globally, and to harmonize and standardize biodiversity data from disparate sources to capture a minimum set of critical variables required to study, report and manage biodiversity change. Here, we assess the challenges of a ‘Big Data’ approach to building global EBV data products across taxa and spatiotemporal scales, focusing on species distribution and abundance. The majority of currently available data on species distributions derives from incidentally reported observations or from surveys where presence-only or presence–absence data are sampled repeatedly with standardized protocols. Most abundance data come from opportunistic population counts or from population time series using standardized protocols (e.g. repeated surveys of the same population from single or multiple sites). Enormous complexity exists in integrating these heterogeneous, multi-source data sets across space, time, taxa and different sampling methods. Integration of such data into global EBV data products requires correcting biases introduced by imperfect detection and varying sampling effort, dealing with different spatial resolution and extents, harmonizing measurement units from different data sources or sampling methods, applying statistical tools and models for spatial inter- or extrapolation, and quantifying sources of uncertainty and errors in data and models. To support the development of EBVs by the Group on Earth Observations Biodiversity Observation Network (GEO BON), we identify 11 key workflow steps that will operationalize the process of building EBV data products within and across research infrastructures worldwide. These workflow steps take multiple sequential activities into account, including identification and aggregation of various raw data sources, data quality control, taxonomic name matching and statistical modelling of integrated data. We illustrate these steps with concrete examples from existing citizen science and professional monitoring projects, including eBird, the Tropical Ecology Assessment and Monitoring network, the Living Planet Index and the Baltic Sea zooplankton monitoring. The identified workflow steps are applicable to both terrestrial and aquatic systems and a broad range of spatial, temporal and taxonomic scales. They depend on clear, findable and accessible metadata, and we provide an overview of current data and metadata standards. Several challenges remain to be solved for building global EBV data products: (i) developing tools and models for combining heterogeneous, multi-source data sets and filling data gaps in geographic, temporal and taxonomic coverage, (ii) integrating emerging methods and technologies for data collection such as citizen science, sensor networks, DNA-based techniques and satellite remote sensing, (iii) solving major technical issues related to data product structure, data storage, execution of workflows and the production process/cycle as well as approaching technical interoperability among research infrastructures, (iv) allowing semantic interoperability by developing and adopting standards and tools for capturing consistent data and metadata, and (v) ensuring legal interoperability by endorsing open data or data that are free from restrictions on use, modification and sharing. Addressing these challenges is critical for biodiversity research and for assessing progress towards conservation policy targets and sustainable development goals., published version, peerReviewed

Data from well-designed experiments provide the strongest evidence of causation in biodiversity studies. However, for many species the collection of these data is not scalable to the spatial and temporal extents required to understand patterns at the population level. Only data collected from citizen science projects can gather sufficient quantities of data, but data collected from volunteers are inherently noisy and heterogeneous. Here we describe a ‘Big Data’ approach to improve the data quality in eBird, a global citizen science project that gathers bird observations. First, eBird’s data submission design ensures that all data meet high standards of completeness and accuracy. Second, we take a ‘sensor calibration’ approach to measure individual variation in eBird participant’s ability to detect and identify birds. Third, we use species distribution models to fill in data gaps. Finally, we provide examples of novel analyses exploring population-level patterns in bird distributions.

Citizen science, big data, and a habitat marketplace enable dynamic habitat for migratory birds in California’s Central Valley., In an era of unprecedented and rapid global change, dynamic conservation strategies that tailor the delivery of habitat to when and where it is most needed can be critical for the persistence of species, especially those with diverse and dispersed habitat requirements. We demonstrate the effectiveness of such a strategy for migratory waterbirds. We analyzed citizen science and satellite data to develop predictive models of bird populations and the availability of wetlands, which we used to determine temporal and spatial gaps in habitat during a vital stage of the annual migration. We then filled those gaps using a reverse auction marketplace to incent qualifying landowners to create temporary wetlands on their properties. This approach is a cost-effective way of adaptively meeting habitat needs for migratory species, optimizes conservation outcomes relative to investment, and can be applied broadly to other conservation challenges.

Automated classification of organisms to species based on their vocalizations would contribute tremendously to abilities to monitor biodiversity, with a wide range of applications in the field of ecology. In particular, automated classification of migrating birds' flight calls could yield new biological insights and conservation applications for birds that vocalize during migration. In this paper we explore state-of-the-art classification techniques for large-vocabulary bird species classification from flight calls. In particular, we contrast a “shallow learning” approach based on unsupervised dictionary learning with a deep convolutional neural network combined with data augmentation. We show that the two models perform comparably on a dataset of 5428 flight calls spanning 43 different species, with both significantly outperforming an MFCC baseline. Finally, we show that by combining the models using a simple late-fusion approach we can further improve the results, obtaining a state-of-the-art classification accuracy of 0.96.

Understanding the susceptibility of highly mobile taxa such as migratory birds to global change requires information on geographic patterns of occurrence across the annual cycle. Neotropical migrants that breed in North America and winter in Central America occur in high concentrations on their non-breeding grounds where they spend the majority of the year and where habitat loss has been associated with population declines. Here, we use eBird data to model weekly patterns of abundance and occurrence for 21 forest passerine species that winter in Central America. We estimate species’ distributional dynamics across the annual cycle, which we use to determine how species are currently associated with public protected areas and projected changes in climate and land-use. The effects of global change on the non-breeding grounds is characterized by decreasing precipitation, especially during the summer, and the conversion of forest to cropland, grassland, or peri-urban. The effects of global change on the breeding grounds are characterized by increasing winter precipitation, higher temperatures, and the conversion of forest to peri-urban. During spring and autumn migration, species are projected to encounter higher temperatures, forests that have been converted to peri-urban, and increased precipitation during spring migration. Based on current distributional dynamics, susceptibility to global change is characterized by the loss of forested habitats on the non-breeding grounds, warming temperatures during migration and on the breeding grounds, and declining summer rainfall on the non-breeding grounds. Public protected areas with low and medium protection status are more prevalent on the non-breeding grounds, suggesting that management opportunities currently exist to mitigate near-term non-breeding habitat losses. These efforts would affect more individuals of more species during a longer period of the annual cycle, which may create additional opportunities for species to respond to changes in habitat or phenology that are likely to develop under climate change.