Introduction

Hyla performed an audio survey of birds in the South Bend area of Conway, Massachusetts over the month of May, 2022, during migration season. We collected 1125 hours of audio across 7 survey sites. We used the AI software BirdNET to analyze the audio for bird calls and created a list of 116 bird species observed across all sites for the month of May. We compared this list to a list compiled by local birdwatchers, and to the 2011 Massachusetts Breeding Bird Atlas, and found substantial agreement.

Study Area

We conducted our study in an area of Conway, Massachusetts colloquially known as the "South Bend," which is a small ridge of hills nearly encircled by a bend in the South River. After obtaining permission from landowners, we placed our recorders across three contiguous plots that, together, form most of the westward facing slope of the South Bend. The area has a vertical elevation span of about 400', stretching from the South River itself all the way up to the top of the ridge at the center of the South Bend. 

Survey area shown in pink. Grey lines partition Massachusetts Breeding Bird Atlas Survey blocks from 2011 and are labelled according to block numbers assigned by atlas coordinators.
Our survey area stretches across blocks /02 (top left), /03 (bottom left), and /06 (bottom right).

As is typical in most of rural Northeastern US, the landscape is second growth northern hardwood forest fragmented with a mosaic of micro-habitats, including forest, wetlands, river banks, meadows, shrublands, and managed grasslands (hayfields).

We placed recorders in all types of habitats found within the study area with the exception of grassland/hayfields, which were excluded to ensure the privacy of landowners.

Recording

A field recorder peeks out of an Elm tree at Site 8, a Phragmites reed marsh at the edge of a Hemlock forest adjacent to a hayfield.

Analysis

We analyzed the audio using the AI software BirdNET, created by the Cornell Lab of Ornithology. BirdNET detects bird calls using a deep neural network. It seems clear to us that BirdNET is an important development in computational bioacoustics, for the following reasons:

• BirdNET is a true AI. A deep artificial neural network model (DNN) designed from ResNet architectures, consisting of 157 layers and over 27 million parameters, BirdNET was trained and validated using multiple datasets and sophisticated augmentation, resulting in a system that can identify 984 species of birds across two continents from almost any audio recording of sufficient quality with very high accuracy and precision.
• BirdNET was designed specifically for wildlife conservation and restoration. Its design is the result of a collaboration between the Cornell Lab of Ornithology, Chemnitz University of Technology, the German Ministry for Education and Research, as well as the online public repositories and citizen science apps Xeno-Canto and eBird, and is released as an open-source package.
• BirdNET can be used by laypersons to conduct high quality estimates of biodiversity. BirdNET is the first open-source AI bioacoustic classifier with an interface that makes it useful to almost anyone who owns a smartphone or computer.

BirdNET labels the input audio with species detected at a given time, including a confidence score. We used the confidence score 0.5 as a cutoff threshold for determining a likely observation, following Kahl 2021. We tabulated a list of 116 species detected across all of our sites, and also created a chart showing the variation in bird call density over the month. The species accumulation curve shows how many species were detected by a certain date. You can browse our raw BirdNET output as well. 

We also gathered two additional species lists collected through manual methods to informally improve our confidence that BirdNET's identifications were accurate within the specific context of our study's local area and time. First, we collaborated with two local birders who made frequent walks through the study area and compiled their observation lists with those of Alex Milne, who also observed several birds while placing and maintaining the recorders. 79 birds were observed by this small community, with 76 of these corresponding to birds also observed by BirdNET. Second, we retrieved site-specific historical data from the 2011 Massachusetts Breeding Bird Atlas survey from three survey blocks that correspond to our study area, for which 120 species were observed by volunteers in 2011. Our survey finds 107 species in common with the Massachusetts Breeding Bird Atlas survey.

Conclusion

While traditional methods using expert human observers - like point count and transect studies - are and will likely continue to be gold standards, they are slow and expensive to conduct. Given the rate and consequences of biodiversity loss, describing biodiversity across the world's ecosystems must be accelerated to a pace faster than the global workforce of trained human experts can likely supply. This is one application where Earth-centered AIs like BirdNET can help. 

This pilot survey demonstrates the potential of AI-enabled bioacoustic monitoring. In the future we plan to collect data at different time points for the same sites, allowing us to track species diversity across time. AI-enabled bioacoustic surveys are an objective, low-impact method of assessing ecosystem diversity, for instance in order to assess the impact of land management practices on ecosystem health. Contact alex@hyla.ai or erika@hyla.ai if you are interested in future collaborations.

Charts and Tables

Table 1 - Descriptive Statistics, All Sites, Pooled

Table 2 - Contextual Community Lists

Species Accumulation Curve (SAC)

Species Accumulation Curves (SACs) show the rate at which species are discovered by a survey over time. Here, a simple SAC curve graphically represents how many species BirdNET’s algorithm observed across 7 sites during the month of May as a single pooled curve.

Species Prescence, All Sites, Pooled

(Confidence >0.5, N = 115)

Appendix