Scripts¶

The code that turns the SmartWilds field archive into the released Darwin Core tables, and the checks that were run on them. The scripts are grouped by stage:

flowchart LR
  A[Field archive<br/>raw media + logs] --> B[darwin_core/<br/>generate tables]
  B --> C[validation/<br/>audit tables]
  B --> D[arrival_traces/<br/>workload streams]
  E[georeferencing/<br/>animal positions] --> B
  A --> F[privacy_screening/<br/>people in media]
  B --> G[(Darwin Core<br/>Event + Occurrence)]

Directory	Stage	Page
`scripts/darwin_core/`	Build the Event and Occurrence tables for every modality and merge them; reconcile raw WAV files against the acoustic Events	Darwin Core generation
`scripts/validation/`	Read-only integrity audit of a generated table set	Validation
`scripts/arrival_traces/`	Campaign-2 arrival streams for the systems analysis	Arrival traces
`scripts/georeferencing/`	Project camera-trap and drone detection boxes to ground coordinates	Georeferencing
`scripts/privacy_screening/`	Find people in the media, triage by identifiability, plan drone trims	Privacy screening
`examples/quickstart.py`	Load and summarize the released tables	Example

Environment variables¶

Every script reads its input locations from environment variables, so nothing depends on the layout of the machine the pipeline originally ran on. Set the ones you need before running a script.

Variable	Used by	Meaning
`SMARTWILDS_ARCHIVE`	all	Root of the field archive: the `thewilds_<campaign>/` directories and `processing_output/`. Default: current directory
`SMARTWILDS_DWC_OUT`	`darwin_core/`	Where the generated Darwin Core tables are written. Default: `<archive>/processing_output/darwin_core_release`
`SMARTWILDS_DETECTIONS`	`darwin_core/generate_darwin_core.py`	External SAM3 + BioCLIP detection workspace (per-video prediction CSVs)
`SMARTWILDS_RELEASE`	`privacy_screening/trim_drones.py`	Root of the staged per-modality datasets
`SMARTWILDS_SCRUBBED_MEDIA`	`privacy_screening/trim_drones.py`	Output directory for trimmed drone videos
`MEGADETECTOR_WEIGHTS`	`privacy_screening/screen_videos.py`	Path to `md_v5a.0.0.pt`
`CAMERA_TRAP_GEOLOCATION_SRC`	`georeferencing/geolocate_camera_trap_treeoflife_predictions.py`	`src/` directory of a checkout of Imageomics/camera-trap-geolocation

Installation¶

pip install -r requirements.txt

The core scripts (darwin_core/, validation/, arrival_traces/) use only the Python standard library. pandas, numpy, pillow, and opencv-python-headless are needed for the example and for privacy screening; MegaDetector, av, and imageio-ffmpeg only for screening (see the commented block in requirements.txt).

Archive layout the scripts expect¶

<archive>/
├── thewilds_20250630_20250703/        # campaign 1
│   ├── camera_traps/<site>-<camera>/<SD card>/...
│   ├── bioacoustics/<site>-<recorder>/...
│   └── drones/<flight>/{metadata,videos}/
├── thewilds_20250804_20250828/        # campaign 2 (+ citizen_science/)
├── thewilds_20250908_20250908/        # campaign 3
└── processing_output/
    ├── detections/                    # detector outputs, timestamp map, geolocation
    └── darwin_core_release/           # SMARTWILDS_DWC_OUT

Packaging the tables into the per-modality Hugging Face datasets and rendering the dataset cards happens in the project's internal release staging and is not part of this repository.