Exploring the Intersection of Ecology and Data Science via Bioacoustics

Hello! I am Joanna, a part-time student from NTU CCDS Masters in Data Science. In this entry, I will be summarizing my recent work for the module SD6106 Capstone Project, which I completed with the supervision of Assoc. Prof. Eleanor Slade and Ms Nicole Dorville of the Tropical Ecology & Entomology Lab (TEE Lab).

My interest in how data science can contribute to ecology led me to a serendipitous encounter at a sharing session by Prof Slade and Dr Chew Li Yuen about their work on sustainability and conservation management in Sabah. Inspired by their work, I reached out to the TEE Lab to explore a collaboration on my capstone project. Like other research fields, ecology has undoubtedly been touched by the data deluge and the rising trend of machine learning. This made it easy to conceive a cross disciplinary data science and ecology project — Evaluation of Bioacoustic Devices for Wildlife Monitoring.

“Bio” refers to (wild)life, while “Acoustics” relates to sound. Bioacoustics is therefore related to the sound of wildlife, which can be studied to tell us more about the dynamics of the ecosystem. Bioacoustic devices like the Song Meter Micro (Wildlife Acoustics), are autonomous recording units that can be deployed for long-term recordings, allowing researchers to easily amass hours of acoustic data which can then be analyzed for insights.

Setting up the Song Meter is quick and easy, requiring just a few clicks on the mobile application. Being small and lightweight, the Song Meter Micro can conveniently be attached to trees/poles in urban parks for recording of wildlife sounds.

As you can imagine, the manual review of hours of recordings can get laborious and time consuming! Luckily for aspiring and sound ecologists of present days, there exists automated tools for analysis of sound data! BirdNET (Cornell Lab of Ornithology) is an example of an automated species classifier for birds that is also behind the eponymously named bird sound ID mobile application. It is able to identify bird species given a sound recording of a bird call. Behind the magic is a convolutional neural network trained on the recordings of thousands of bird species (mostly from North America and Europe), which allows the classifier to predict bird species from 3-second sound segments with a confidence score.

BirdNET: How it works. (Source: https://birdnet.cornell.edu)

My project aimed to investigate the use of Song Meter Micro in an urban environment, as urban areas are typically noisier than natural environments, which are more commonly reported in the literature. Concurrently, I assessed the performance of BirdNET in capturing the bird fauna of Singapore by comparing its predictions with point count survey results. The overarching goal of the project was to identify and define the parameters/tools suitable for bioacoustics research in Singapore using light-weight acoustic devices like Song Meter Micro.

My project was comprised of two parts: Distance Estimation and Observation, with the following research questions:

Part 1. Distance estimation

• Does distance from the sound source to the Song Meter Micro acoustic recording device affect detectability of animal calls on the spectrogram by humans?

• Does ambient noise affect detectability of animal calls on the spectrogram by humans?

Part 2. Observation

• How well does the automated acoustic classifier BirdNET capture the bird fauna of Singapore?

Part 1: Distance estimation

The distance estimation experiments were conducted in 3 sites: Mandai, Punggol, and Pioneer, representing increasing ambient noise levels. A portable Bluetooth speaker broadcasting a bird/cricket call (downloaded from Xeno-canto) was sequentially moved further away from the Song Meter, which continuously recorded sounds during the session. After some analysis (data wrangling, randomization and visual and aural annotation by volunteers), the data was used to build a generalised linear model to model the detection probability of the broadcasted animal calls.

Results of distance estimation experiments. Interestingly, we are more likely to detect cricket chirps than calls of the “uwu” bird, the Asian Koel, at the same distances in sound recordings.

The most striking observation was the difference in shapes of the curves across different animal types, which could be due to the masking of the lower frequency Koel call by traffic noises. The distinction across study sites were also noteworthy. For instance, I clearly recall an instance during the Mandai recording session, when a cricket seemed to be competing (in terms of volume) with my broadcasted cricket call. This was likely the reason for the unique shape of the cricket curve observed at Mandai. It also highlighted the difficulties of conducting controlled experiments in outdoor field environments.

Part 2: Observation

Data comprising approximately 1 month of acoustic recording (1 min every 5 min) was collected (by TEE lab members) from Song Meter Micros deployed in six study sites: Bishan-Ang Mo Kio Park (BAMK), Bukit Timah Nature Reserve (BTNR), Jurong Lakeside Garden (JLG), MacRitchie Reservoir Park (McRitchie), Singapore Botanic Gardens (SBG), Sungei Buloh Wetland Reserve (SBWR). The recordings were analyzed with BirdNET, and the resulting predictions and their confidence scores were compared with presence-absence data from Garden Bird Watch, a citizen science survey. This analysis involved several rounds of data curation to align species databases from various sources. The results of the comparison were:

Bukit Timah Nature Reserve (BTNR) consistently has the highest false negative rate across all confidence cutoffs. Sites that were more urban, like Bishan-Ang Mo Kio Park (BAMK) and Jurong Lake Gardens (JLG), had lower false negative rates.

The result suggested that false negative rate was higher in areas with dense vegetation. It is possible that tropical bird species found in such environments, like in BTNR, were underrepresented in BirdNET, which was mainly trained on North American and European birds. I also attempted to manually validate the BirdNET predictions of five bird species with high confidence scores by comparing the recordings with high quality recordings of the same species from Xeno-canto.

Example of incorrect and correct BirdNET predictions. Left panels show the screen grab of Raven Lite, the software used to visualize sound recordings. Right panels show the reference recordings from Xeno-canto

Due to my limited expertise in bird call recognition, I found this process to be quite time consuming, as I had to repeatedly switch between listening to the BirdNET prediction and reference recordings. Overall, I felt that the manual audio validation was fairly subjective, though this would be less of an issue if I were a bird call expert. Interestingly, the results hinted at the possibility of geographical training bias:

Among the two well-sampled bird species, the Arctic Warbler, a migratory bird from North America, had a low error rate, whereas the Stork-billed Kingfisher, a bird only found in Southeast Asia, had a high error rate.

Unfortunately, as bird species like Indian Paradise-Flycatcher and Tiger Shrike were not sufficiently sampled in this data set, I was unable to deduce much from their error rates.

Summary of this study

High levels of anthropogenic noise in urban areas like Singapore limits the detection range of acoustic recording devices and contributes to additional complexity in bioacoustics studies. This study is one of the first to provide additional insights to the site characteristics and parameters that should be considered before long term deployment in urban environments.

Reflections

Having spent most of my working hours indoors in research labs or behind a computer, one of the highlights for me was the chance to engage in outdoor fieldwork. Although my capstone project was computational, I was involved in the distance estimation experiments, which required broadcasting and recording of animal calls (at a noticeable volume) in public areas during weekend mornings and evenings. While the sessions were brief and well within recommended sound limits, my experiments understandably drew some disapproving glances from passersby.

In one instance, a member of the public lamented, “School projects should be done in schools”, after inquiring about the nature of my activities. As a small garden city, encounters with the public are inevitable when conducting wildlife research in Singapore. This experience has made me more empathetic towards research activities that may present as a transient inconvenience in public spaces.

Setting up for the distance estimation experiment after my prayers for the rain to stop were answered. Ironically, fieldwork in Singapore may not be very field-like! Photo credit: Nicole 2024.

Balancing a full-time job, evening classes, ongoing quizzes and experiments/analysis/literature reading for this project was no easy task. Given these challenges, I am especially grateful for the invaluable guidance and support I received from members of the TEE lab throughout this experience.

Thank you, Prof Eleanor, for your positive feedback during our discussions as well as detailed critique of my report! Thank you, Nicole, for you constant encouragement and honest insights about your experience in ecology research! Thank you, Xin Rui, for your patience when delivering the crash course on stats! Thank you, TEE Lab, for this enriching experience!