# Building a Reliable Citizen-Science Bird Survey

Citizen-science bird surveys can cover far more ground than professional field teams, but volunteer observations arrive with uneven effort, experience, and equipment. Our spring survey addressed this by assigning participants fixed 1 km grid cells and asking them to complete three ten-minute point counts between sunrise and 10:00 a.m. Each checklist recorded detected species, counts, start time, weather, and whether the observer identified birds by sight or sound.

## Capturing Consistent Field Data

The mobile form used controlled species names, numeric validation, and required effort metadata. GPS coordinates were captured automatically but rounded before publication to protect sensitive nesting sites. The survey configuration was versioned alongside the analysis code so that changes to count duration or weather thresholds remained traceable.

```yaml
protocol:
  duration_minutes: 10
  visits_per_site: 3
  max_wind_beaufort: 3
  include_flyovers: false
```

## Quality Control

Automated checks flagged records outside assigned cells, counts submitted after the survey window, and unusually high species totals. Flags triggered review rather than automatic deletion: a flock of 600 migrating blackbirds may be valid, while a single accidental extra zero is not. Regional coordinators compared questionable observations with field notes, photographs, audio recordings, and nearby checklists.

## Accounting for Unequal Effort

Raw totals can be misleading because detection varies with weather, habitat, time, and observer skill. We therefore modeled species occurrence using site and visit as separate levels, with wind speed, start time, and observer experience included as covariates. Repeated visits helped distinguish a bird’s probable absence from a failure to detect it during one count.

## Results and Reproducibility

Across 184 grid cells, 127 volunteers submitted 1,492 complete point counts and documented 163 species. Woodland specialists were most strongly associated with cells containing larger connected forest patches, while several open-country species appeared less frequently near expanding residential areas. Because the project was observational, these patterns were reported as associations rather than evidence of causation.

## Lessons for the Next Season

The largest improvement came from clearer protocol design, not more sophisticated modeling. Short training recordings reduced common song-identification errors, and immediate form validation prevented many incomplete submissions. Next season, we will add offline map support, rotate a subset of sites among observers to estimate observer effects, and publish anonymized data with a reproducible analysis notebook.