Independent Research — Brooklyn, NY

Automated avian monitoring for urban ecosystems

A low-cost Raspberry Pi machine learning system continuously detecting and logging bird species in urban Brooklyn — testing whether automated systems can scale urban biodiversity monitoring.

Research Question

Can a low-cost, automated machine learning detection system produce species diversity data statistically consistent with Cornell Lab of Ornithology's Project FeederWatch citizen science dataset — and could such systems scale urban biodiversity monitoring beyond what citizen science alone can cover?

Louis Davies·Brooklyn, NY·2026
§1

Project goals

This project sits at the intersection of machine learning, field ecology, and urban science. The immediate goal is a single-site deployment that generates a reliable, continuous dataset of automated bird detections through a full season.

The longer-term goal is comparative: to test whether that automated dataset is statistically consistent with Project FeederWatch's citizen-science observations for the same region, and to use that comparison to evaluate whether low-cost automated systems could extend the reach of urban biodiversity monitoring beyond what volunteer observers alone can cover.

Active
The model, up and running
Detection system deployed and continuously classifying visiting birds in real time in Brooklyn.
Upcoming
Publishing first research & findings
Two initial write-ups:
  • Comparing automated detections against Project FeederWatch citizen-science data
  • Analyzing correlations between detection activity and weather patterns
Future
Research paper
A formal write-up of the full season's methodology, results, and analysis.
Future
Submitting to competitions
Entering the completed research into science competitions, including the Regeneron Science Talent Search.
§2

How the system works

The detection system runs continuously on a Raspberry Pi paired with a fixed camera module, classifying visiting birds in real time and logging each detection with species, confidence score, and timestamp.

HARDWARE
Raspberry Pi
Runs the classifier on-device, continuously, with no cloud dependency for inference.
MODEL
EfficientNet B0
10-class image classifier, 60% confidence threshold before a detection is logged.
CAMERA
Fixed-angle module
Camera module locks into place with a fixed angle, enabling continuous outdoor deployment through rain and storms.
LOGGING
Species · Confidence · Timestamp
Each detection is written to a running log for later comparison against Project FeederWatch data.
Note — the enclosure
The Pi, camera, and wiring sit inside a custom enclosure I designed in Fusion 360 and 3D printed, rather than an off-the-shelf case. That matters for a few reasons:
  • Weatherproofing — the electronics need to survive rain and temperature swings through a full outdoor deployment season, not just a single afternoon.
  • Fixed camera geometry — the mount holds the camera at a precise, unchanging angle. Since the classifier is trained on images from that exact framing, any drift in camera position would introduce noise the model was never trained on.
  • Ventilation — vents are built into the design to prevent the Pi from overheating during continuous operation.
  • Iteration — designing it myself meant I could revise the fit after early field tests, rather than adapting the whole setup around a generic case.
Camera enclosure and bird feeder deployed side by side in the backyard
Deployed setup with feeder, casing, mount, camera module, and support strings labeled
Note — species selection
10-class model deliberately limited to common Brooklyn feeder species. Narrower scope improves per-class accuracy versus attempting to classify rare visitors with insufficient training data.
Target species — banding codes
AMGOAmerican Goldfinch
BCCHBlack-capped Chickadee
BLJABlue Jay
DEJUDark-eyed Junco
DOWODowny Woodpecker
HOFIHouse Finch
HOSPHouse Sparrow
MODOMourning Dove
NOCANorthern Cardinal
TUTITufted Titmouse
§3

Detection log & analysis

Detection logs, R analysis, and comparative Project FeederWatch visualizations will be published here as the dataset grows. Data is updated weekly. The full raw dataset will be made available for download once sufficient records have been collected.

[status] data collection underway
[next update] summer 2026
[waiting for records]
§4

About this project

I'm a rising junior at Packer Collegiate Institute in Brooklyn, NY, interested in the intersection of data science, machine learning, and environmental science — specifically how computational tools can help us understand and protect urban ecosystems.

NYC FeederWatch AI began in February 2026 as an independent research project. The core question driving it: can a low-cost automated system produce biodiversity data as reliable as human citizen science? The answer has implications for how conservation organizations monitor urban wildlife at scale.

§5

Get in touch

I welcome correspondence from researchers, conservationists, and anyone interested in automated biodiversity monitoring.