Skip to content

Data overview

SmartWilds has four event-bearing modalities, co-located across overlapping sites and weeks, plus field-note and text context streams. Every modality is expressed in the same Darwin Core schema and joined into one Event hierarchy rooted at TheWilds_2025.

TheWilds_2025 (study, 2025-06-30/2025-09-08)
├── SmartWilds_camera_traps_curated (Project)
│   └── SensorDeployment
│       └── Image
└── Other modality Projects
    └── SensorDeployment / flight session / observation session
        └── AudioRecording | MediaEvent
Dataset Content
smartwilds-camera-traps 235,778 processed images, 19,785 Events, 72,775 operational Occurrences
smartwilds-bioacoustics 2,371 WAV recordings, 9,960 Occurrences
smartwilds-drones 39 videos, 15 flight sessions, 318,215 Occurrences
smartwilds-citizen-science 1,103 media files, 622 Occurrences
smartwilds (study level) merged Event table, text records, sensor configurations, arrival traces

The current camera-trap tables are described in the camera-trap release note. Counts, inventories, and every repair are in the supplemental material; the defects themselves are in the defect inventory.

Using the data correctly

Pick a verification floor before analysis

Almost every Occurrence is a machine label, and the status vocabulary differs by table. Camera-trap labels carry a labelProvenance (triage_resolved, triage_reviewed, triage_forced_1nn, human-edit, human-reviewed); drone and acoustic labels are unverified or validated; the 12 human-added drone starling boxes are human-supported; limited-resolution. The drone labels come from BioCLIP 2 (the TreeOfLife-200M model, imageomics/bioclip-2) run in its open-world Tree of Life mode, which predicts a taxon with its full rank hierarchy rather than choosing among a fixed label list; species-rank accuracy on human-annotated subsets is 0.271 and class-rank accuracy is 0.845 (supplement, S9).

  • Occurrences are detection crops, not individuals. Bursts and adjacent drone frames repeat the same animal.
  • Use the corrected timestamps. Camera clocks were wrong on most deployments, and burned-in overlays show the raw clock. Every image carries its correction method and confidence (images.csv).
  • A zero occurrence_count is not evidence of absence.
  • Report your choices. State the unit of observation, the verification floor, and the timestamp policy alongside any derived metric.

Where the files are