About this station

This is an automated backyard bird feeder station. A camera and a microphone watch a feeder; the video and audio are processed locally, and what you see here is what the station made of it: which birds came, when, and how sure it is. Every species claim carries its identification tier, and the daily narrative is written from the counts alone.

Location: the published location is randomized within 1 km of the station.

What is never published

• Only birds are published. People, vehicles, and faces are never stored; wild mammals the nightly pass finds are kept in the station's private records and never appear here.
• Raw video is a short rolling buffer. Only crops and short clips of detected birds and wild mammals are kept past it, and only bird crops are published.
• No individual identity is tracked for non-target subjects, ever.
• Nothing here describes when this household is occupied: the published location is randomized within 1 km of the station, and no uptime, health, or camera schedule is published. Times on this site are when a bird was seen or heard.

Reading the labels

• No badge: the station is confident of the species.
• Probable: likely, but the classifier was not sure enough to say so plainly.
• Possible: several candidates; the names listed are the top guesses, best first.
• Unverified: no second line of evidence yet. Either an unusual species for this area seen without corroboration, or a species the microphone picked up only a few times. Treat it as a maybe; most single detections of an out-of-place species are classifier errors.

The hardware and software

Two devices on the porch capture the video and the audio and send them over Wi-Fi to a workstation indoors, which does all of the identification. Nothing is uploaded to a cloud service or a third-party API along the way, and only the finished pages leave the house.

A TP-Link Tapo C120 camera, about six feet from the feeders, supplies the video. The audio comes from a dedicated microphone: a lavalier mic hanging in the open on the porch under a furry windscreen, wired to a Raspberry Pi 3 B+ in a weatherproof box that records it at 48 kHz through a USB sound adapter, filters out electrical hum, and streams it over Wi-Fi. The dedicated microphone replaced the camera's built-in one on September 13, 2026; the camera's audio is telephone quality (8 kHz), which cuts off much of the range songbirds sing in.

Frigate watches the video stream and tracks birds with an exported D-FINE object detector running on an Nvidia RTX 3090, producing the crops and clips the rest of the pipeline works from. Each night a second pass runs the same detector over the day's motion-triggered recordings at full resolution to find birds the live pass missed, hummingbirds especially; its finds stay off this site until a person has checked them. The same nightly pass looks for wild mammals such as squirrels and chipmunks, which are kept for the station's own records and not published. BioCLIP 2, a fine-grained species classifier, and Frigate's own built-in classifier each get a vote on a crop, checked against a regional species list; a person's review can override either. BirdNET-Go listens to the audio independently and never sees the video. A local vision-language model, qwen3.8:27b served by Ollama, looks at the best crop of a visit to describe it and help settle disagreements between the other two. It also writes the recap on the Log page, once a day, shortly before the next morning's sunrise, from text only: the day's computed counts and timeline, the week before it for comparison, and the station's own log of changes such as feeder refills or new hardware. The database, the review dashboard, and this public site are all built by the same private project (Fieldstation) running on that same workstation.

The hourly weather is measured by a nearby station of the North Carolina ECONet; cloud cover, which it does not measure, comes from the Open-Meteo weather model. Weather data provided by the North Carolina State Climate Office.

How the station decides

Two separate systems, with different ways of being wrong. A microphone runs an acoustic classifier (BirdNET) over three-second windows of sound. A camera runs an object detector, then an image classifier on crops of whatever it tracked, and a vision-language model describes the best crops. None of them knows what the others found, so "seen" and "heard" are not comparable counts: a wren calling from the hedge all morning is heard dozens of times and seen never; a titmouse that visits the feeder in silence is the reverse. A species with two hundred records heard and one seen tells you about the sensors, not the bird.

The microphone first recorded a bird on September 7, 2026; the camera first recorded a visit on September 9, 2026. Days before the camera's first visit are audio only, and a jump in visits on that day is the hardware arriving, not the birds. The same goes for sound on September 13, 2026, when the dedicated microphone replaced the camera's: it records several times as many detections a day, so heard counts before and after that date do not compare.

Known failure modes. The acoustic classifier mistakes distant traffic, squirrels, insects and overlapping song for birds, and one loud call can score as two or three detections in the same minute. Its confidence is not calibrated across species: a wrong shorebird has scored 0.96 here while a real Blue Jay averages about 0.6, so this site does not rank claims by confidence. Camera crops are small, often little more than a hundred pixels across, well below what fine-grained bird classifiers were trained on, so the image classifier confuses similar shapes and now and then names a species that has never been in this county. Rare species with a single detection are usually errors.

What makes a species count. A species is confirmed here when a visual record passes the publication check (an identification that is either expected for this area or backed by a second line of evidence: another confident visit, an overlapping call, the vision model naming the same bird from its field marks, or a person's review), or when the microphone has recorded it at least 5 times across at least 2 separate hours. Everything else is unverified: still listed, still in the raw detection totals, and framed as the classifier's output rather than the station's claim. A highlight's caption is shown only when the vision model named the same species the classifier did. Verify and reject decisions made by a person apply retroactively: a rejected visit leaves its day and the day's write-ups are regenerated.

Reading the counts. A day's "species" figure counts seen and heard species once each. "Feeder visits" counts camera visits after consecutive detections of one bird are grouped. "Audio detections" counts three-second classifier hits, so one bird calling for a minute produces many. The hour grid keeps seen and heard in separate columns for the same reason.