SmartWilds — supplemental material¶

Camera-trap update (2026-11-08). The legacy camera counts, Event hierarchy, first-generation Tree-of-Life method statements, and camera detection-arrival rows in this historical supplement do not describe the current camera release. The canonical camera release has 72,775 original-SAM3-geometry Occurrences, 19,785 camera Events, a 235,778-image inventory, and 70,466 Campaign-2 operational-inference arrivals. Its combined multimodal table has 401,602 Occurrences and 22,242 Events. See CAMERA_TRAP_RELEASE.md for the release-facing scope and provenance. The historical sections remain for auditability and must not be used as current camera-release statistics.

Supplemental material for the SmartWilds datasets. This contains the full data inventories details, the timestamp-repair methods, the standards crosswalk, the field dictionary, the arrival statistics, and the complete defect inventory.

Last updated: 2026-11-10. Release: SmartWilds release build 2026-10-22, camera tables updated 2026-11-08. It supersedes the 2026-07-24 integration pass. Paths such as processing_output/... are relative to the project archive; the released files are in the per-modality datasets of the SmartWilds collection.

S1. Extended glossary¶

Term	Source	Meaning in this release	What it is not
Event	Darwin Core	A sampling act with a date, location, protocol, and `parentEventID`. Five levels in use: `Project`, `SensorDeployment`, `AudioRecording`, `MediaEvent`, `ObservationSession`.	Not an arrival in a stream. One Event can contain 10⁵ arrivals.
Occurrence	Darwin Core	One organism record at a place and time, linked by `eventID`. Usually one detection crop.	Not an individual animal. Adjacent drone frames yield many Occurrences of the same animal.
`eventType`	Darwin Core	The level of the hierarchy a row occupies.	Not the modality; modality is carried by `parentEventID` and `modality_type`.
`basisOfRecord`	Darwin Core	`MachineObservation` for every detector-derived row; `HumanObservation` for citizen-science and text records and for the 12 human-added drone starling boxes.	Not a quality statement. See `identificationVerificationStatus`.
`identificationVerificationStatus`	Darwin Core	Vocabulary differs by table. Post-triage camera traps: `human-reviewed`, `machine-resolved`. Drones and bioacoustics: `unverified`, `validated`, plus `human-supported; limited-resolution` for the drone starling supplement. Citizen science and text: `verified by expert`. Counts in S9.3.	Not a confidence score. `machine-resolved` and `unverified` are both machine labels; a `validated` or `human-reviewed` record was reviewed by a person.
`dataGeneralizations`	Darwin Core	States whether a coordinate is an observed position, a sensor position, a projected estimate, or a platform-telemetry fallback.	Not a precision estimate; see `coordinateUncertaintyInMeters`.
`informationWithheld`	Darwin Core	Records where locality precision was reduced for sensitive taxa.	—
Sampling effort	Humboldt ext.	Acquisition quantity actually expended, per modality: camera observation-days, recorder-hours, flight-minutes, visitor sessions.	Not comparable across modalities. Not a detection count.
Arrival	systems	One item entering a processing stream. Two layers measured: acquisition (image capture, WAV file) and detection (an Occurrence).	Not a Darwin Core Event.
Deployment	systems	One physical installation of one device at one site over one interval.	Not a device. One device yields several deployments across campaigns.
Duty cycle	systems	Fraction of wall-clock time a device was scheduled to acquire.	Not uptime. A device can be on-schedule and dead.
Detection	systems	A model output: a SAM3 box, a BirdNET hit.	Not an identification.
Identification	Darwin Core	A taxonomic assignment with a stated verification status.	Not implied by a detection.
`locationID`	Darwin Core	Stable identifier for a fixed sensor position or a drone session origin. Null for locality-only records.	Not a comparable spatial sampling unit across modalities.
Coverage	systems	Distinct (date, hour) pairs containing ≥1 Occurrence, per modality.	Not area surveyed. Not detection probability.
Campaign	this release	One field deployment interval. Three: C1 Jun 30–Jul 3, C2 Aug 4–28, C3 Sep 8.	Not a Darwin Core term.

S2. Camera-trap deployment inventory¶

Four camera models, HFOV 55–120°. Configuration varied in capture mode, resolution, burst length, PIR sensitivity, and inter-trigger delay. Every one of these is a workload multiplier (S2.3).

S2.1 Campaign 1 — Jun 30 to Jul 3, 2025¶

Deployment	Site	Model	HFOV	Facing	Mode	Photo	Video	PIR	Images	Notes
TW01-CT02	Nomad Ridge E	GardePro T5NG	56.5°	SE (135°)	Hybrid	4 MP, 3-burst	1080p 10 s MP4+audio	High	7,811	Hourly timelapse, info strip on
TW02-CT03	Nomad Ridge W	GardePro A3	63°	SW (225°)	Photo+Video	8 MP, 3-series	1080p 10 s AVI+audio	High	39	Early battery death; 55 files total on card
TW03-CT01	Pasture D Feeder	GardePro T5NG	56.5°	S (180°)	—	—	—	—	6,661	Near salt lick; config not documented
TW04-CT04	Lake Trail	Trail Camera G600	120°	NW (315°)	Photo+Video	8 MP, 5-burst	1080p 10 s MP4+audio	High	570	Images inverted (mount orientation)

S2.2 Campaign 2 — Aug 4 to Aug 28, 2025¶

Deployment	Site	Model	HFOV	Facing	Mode	Photo	Video	Images	Notes
TW01-CT02	Nomad Ridge E	GardePro T5NG	56.5°	SE (135°)	Photo only	4 MP, 3-burst	—	26,573	Removed Aug 11, relocated to dock
TW02-CT01	Nomad Ridge W	GardePro T5NG	56.5°	SW (225°)	Photo only	4 MP, 3-burst	—	61,217	Clock error; battery died Aug 18–25; 4 card clock states
TW04-CT03	Lake Trail	GardePro A3	63°	NW (315°)	Photo+Video	8 MP, 3-burst	1080p 10 s	984	Reset Aug 7; no data after Aug 25
TW06-CT04	Zebra Shelter	Trail Camera G600	120°	NW (315°)	Photo+Video	8 MP, 5-burst	1080p 10 s	40,907	Battery reported 100% at retrieval (unusual)
TW06-CT05	Zebra Shelter (in)	Reveal Pro 3.0	55°	N (0°)	Photo+Video	4K	720p 10 s	84	Late deploy Aug 25–28
TW07-CT05	Zipline Tower	Reveal Pro 3.0	55°	N (0°)	Photo+Video	4K	720p 10 s	withdrawn	People-only media were not uploaded; metadata only retained; deletion status unknown
TW09-CT02	Dock (left)	GardePro T5NG	56.5°	S (180°)	Photo only	4 MP, 3-burst	—	90,834	Largest deployment (38.5% of all images); relocated Aug 18; battery gap Aug 18–25
TW09-CT05	Dock (right)	Reveal Pro 3.0	55°	SE (135°)	Photo+Video	4K	720p 10 s	9	Detection issues
TW11-CT06	Fence / lake	Reveal Pro 3.0	55°	NE (45°)	Photo+Video	4K	1080p 15 s	89	PIR level 9, 20 s delay, 1 bar cell service

Totals: 235,778 JPG + 5,113 video clips, 260 GB. 13 deployments, 12 retained.

S2.3 Configuration as workload multiplier¶

Parameter	Range in deployment	Workload effect
Burst length	3 or 5 frames/trigger	Direct 3× or 5× multiplier on raw images per ecological event
Capture mode	photo-only / photo+video / hybrid	Video adds a 10–15 s clip per trigger: large storage, frame-extraction compute
Resolution	4 MP / 8 MP / 4K	Bytes and per-image inference time
PIR sensitivity	High, Level 9, motion-detect	Higher sensitivity → more triggers and more blanks
Inter-trigger delay	0 / 10 / 20 s	Longer delay suppresses repeat triggers, lowering effective arrival rate
HFOV	55°–120°	Wide-angle covers more area at lower pixel density per animal
Cellular uplink	Reveal Pro 3.0 only	The only device in the deployment that could participate in near-real-time coordination
Timelapse	TW01-CT02 C1 only	Adds a scheduled arrival stream independent of PIR

S3. Acoustic recorder inventory¶

All units: Wildlife Acoustics Song Meter Micro 2, 48 kHz. 10 deployments across 6 sites. Duty cycle, not deployment duration, determines volume: uncompressed 48 kHz stereo WAV ≈ 33 MB/minute.

Deployment	Site	GPS	Schedule	Duty cycle	Raw WAV	Events	GB
C1
TW05-SM01	Lake Trail Gate O	39.83537, −81.73842	5-min clips hourly (config error — intended 1 hr sunrise + sunset)	8.3%	144	144	—
TW06-SM03	Zebra Shelter	39.83518, −81.73087	5 min/hr, 24/7	8.3%	77	77	—
TW07-SM02	Zipline Tower	39.83288, −81.73455	1 hr sunrise + 1 hr sunset	~8%	12	12	—
TW08-SM04	Fence w/ vegetation	39.82900, −81.74378	5 min/hr, 24/7	8.3%	78	78	—
					311	311	6.0
C2
TW02-SM04	Nomad Ridge Shelter	39.82957, −81.73295	5 min/hr	8.3%	659	581	26.1
TW05-SM06	Lake Trail Gate O	39.82866, −81.73758	sunrise–sunset	~50–60%	100	96	16.6
TW06-SM05	Zebra Shelter	39.82870, −81.73756	15 min/hr	25%	508	508	21.8
TW07-SM02	Zipline Tower	39.82891, −81.74392	sunrise–sunset	~50–60%	339	339	56.4
TW08-SM03	Fence w/ vegetation	39.82892, −81.74392	sunrise–sunset	~50–60%	341	341	56.7
TW09-SM01	Dock	39.83418, −81.73783	sunrise–sunset (deployed Aug 14)	~50–60%	195	195	32.4
					2,142	2,060	210
Total					2,453	2,371	216

Raw-WAV reconciliation. Of the 2,453 WAV paths on disk, 2,371 are AudioRecording Events: 2,307 with BirdNET detections and 64 on which BirdNET ran but detected nothing (Events with no Occurrences). The other 82 are not recordings of their own: 78 are Campaign-1 files copied into the Campaign-2 TW02-SM04 directory during transfer, and 4 TW05-SM06 files produced no BirdNET result (three 3.9–22.0 s fragments and one unreadable file). Every path is classified in data/recordings.csv of the bioacoustics dataset.

Note that TW02-SM04 produced more than twice the files of TW07-SM02 but less than half the bytes — the clearest illustration that duty cycle and file count are independent axes.

S4. Drone platform and session inventory¶

Platform	Campaigns	Video	Telemetry format	Sessions
Parrot Anafi 4K	C1 (Jul 1, 3), C2 (Aug 11, 18, 25)	4K MP4	Parrot JSON / GUTMA DX	10 (+3 telemetry-only)
DJI (unspecified)	C2 (Aug 11)	1080p MOV	CSV, GPS + battery, 10 Hz	1
DJI Air 2	C3 (Sep 8)	4K MP4	Airdata CSV	4

39 raw videos and 39 MediaEvent rows; one MediaEvent (VID_001) has zero Occurrences because it was intentionally excluded after review. 15 video-bearing SensorDeployment flight-session Events. Three more Aug 25 Parrot directories hold telemetry but no video and are not sampling Events. 90 GB total.

Three platforms produced three telemetry schemas with different field names, units (feet vs. metres for altitude), and gimbal-angle conventions. Fields consumed by the projection script, with their per-platform sources:

platform GPS — latitude_y / longitude_y
AGL height — height_sonar(feet) when available, else altitude (m) / altitude_m
heading — gimbal_heading(degrees) or compass_heading(degrees)
pitch — gimbal_pitch(degrees), with a configurable fallback when near-level
FOV — approximate per-platform constant

S5. Camera-trap timestamp repair, in full¶

235,778 images; 0 unmapped; 0 flagged for review. 18 correction methods.

S5.1 Raw timestamp source¶

Source	Images	Share
Filename	193,096	81.9%
EXIF	42,500	18.0%
OCR of burned-in bottom overlay	179	0.08%
Manual read of bottom overlay	3	0.001%

S5.2 Final confidence¶

Confidence	Images	Share	Meaning
high	105,488	44.7%	Verified correct, or anchored to a logged service event
medium	130,290	55.3%	Anchored interpolation assuming a constant clock rate — untestable without a second anchor

S5.3 All 18 correction methods¶

Method	Images	Pattern
`aug_tw09_ct02_sd23_correct_after_move`	59,568	verified correct
`aug_tw09_ct02_sd22_start_anchor`	31,266	per-card anchor
`aug_tw06_ct04_sd02_start_anchor`	27,734	per-card anchor
`aug_tw01_2025_clock_offset`	26,558	constant offset
`aug_tw02_actual_august_clock`	17,301	verified correct
`aug_tw02_sd01_sep_clock_anchor`	17,262	per-card anchor
`aug_tw02_sd01_2023_clock_anchor`	13,689	constant offset (clock reporting 2023)
`aug_tw06_ct04_initial_start_anchor`	13,173	per-card anchor
`aug_tw02_sd11_sep_clock_anchor`	12,965	per-card anchor
`june_tw01_correct`	7,811	verified correct
`june_tw03_correct`	6,661	verified correct
`aug_tw04_sd02_start_anchor`	696	per-card anchor
`june_tw04_start_anchor_inverted_images`	570	per-card anchor + inverted mount
`aug_tw04_sd03_start_anchor`	288	per-card anchor
`bottom_overlay_ocr`	179	OCR of in-frame timestamp
`june_tw02_start_anchor`	39	per-card anchor
`aug_tw01_2023_tail_clock_offset`	15	constant offset (2023 tail)
`manual_bottom_overlay_read`	3	human read

Only 14,472 images (6.1%) were accepted as recorded (june_tw01_correct + june_tw03_correct). The remaining 221,306 (93.9%) required reconstruction.

S5.4 Per-deployment repair summary¶

Dataset	Deployment	Location	Images	First corrected	Last corrected	Methods
C1	TW01-CT02	Nomad Ridge E	7,811	2025-06-30 12:18	2025-07-02 06:50	1
C1	TW02-CT03	Nomad Ridge W	39	2025-06-30 12:34	2025-07-01 00:34	1
C1	TW03-CT01	Pasture D Feeder	6,661	2025-06-30 13:06	2025-07-02 01:04	1
C1	TW04-CT04	Lake Trail	570	2025-06-30 13:25	2025-07-03 16:56	1
C2	TW01-CT02	Nomad Ridge E	26,573	2025-08-04 09:01	2025-08-10 10:15	2
C2	TW02-CT01	Nomad Ridge W	61,217	2025-08-04 09:05	2025-08-18 09:07	4
C2	TW04-CT03	Lake Trail	984	2025-08-07 10:38	2025-08-25 10:49	2
C2	TW06-CT04	Zebra Shelter	40,907	2025-08-04 09:52	2025-08-26 15:35	2
C2	TW06-CT05	Zebra Shelter	84	2025-08-25 15:09	2025-08-28 08:17	2
C2	TW09-CT05	Dock / Waterfront	9	2025-08-11 10:18	2025-08-13 18:20	1
C2	TW09-CT02	Dock, then Nomad Ridge W after Aug 18	90,834	2025-08-11 10:27	2025-08-28 09:44	2
C2	TW11-CT06	Nomad Ridge W	89	2025-08-18 09:00	2025-08-28 02:47	2

Files: processing_output/detections/drone_camera_analysis/camera_time_mapping/ — camera_image_timestamp_map.csv (per-image, 17 columns), camera_timestamp_summary_by_camera.csv, camera_timestamp_summary_by_date.csv, camera_trap_timestamp_corrections.csv, plus .before_bottom_ocr and .before_ocr_interpolation stages.

S6. Drone ground projection¶

The campaigns were flown for population estimation and behavior, not geolocation. Platform GPS localizes a detection to a known pasture, which is the operative spatial reference. Per-animal projection is an optional refinement provided for completeness; no result in any of the three papers depends on it. Monocular distance estimation from a moving platform to moving subjects is unreliable and is not established practice in aerial wildlife survey.

Status	Detections	Share
`ok` (projected position available)	247,340	77.9%
`invalid_altitude`	27,306	8.6%
`missing_drone_gps`	24,467	7.7%
`horizon_ray` (camera near level)	18,324	5.8%
`out_of_range`	122	0.04%

Method: convert the bbox bottom-center pixel to a camera ray, rotate by heading/pitch, intersect with a flat ground plane, project the offset from platform GPS. Dominant error sources: gimbal-pitch convention, terrain relief, altitude source, video stabilisation, FOV assumption, and whether bbox bottom-center is the animal's ground-contact point.

Camera-trap coordinates use the pinhole method from https://github.com/Imageomics/camera-trap-geolocation: bbox height → range, bbox horizontal offset → bearing, projected from the camera GPS. Assumes flat ground and a uniform 0.9 m target height — the dominant error term at a site holding animals from geese to Père David's deer.

Known analysis hazard. An internal audit found a spatial figure that plotted projected animal positions and platform-telemetry positions together without distinguishing them: 15.66% of 292,106 plotted drone points were platform coordinates. The status field was present and correct; the script did not read it. The release carries that status on every drone Occurrence (drone_georef_status in occurrenceRemarks, georef_status in detections.csv); any analysis of drone coordinates must filter on it, and examples/quickstart.py shows the pattern. No automated check of downstream analyses ships with the release.

S7. Standards crosswalk, in full¶

Required information	Standard	Status
Taxon, rank hierarchy, vernacular name	Darwin Core	reused directly
`eventDate`, `eventTime`, event hierarchy	Darwin Core	reused directly
`decimalLatitude`/`Longitude`, `coordinateUncertaintyInMeters`, `geodeticDatum`	Darwin Core	reused directly
`basisOfRecord`, `identificationVerificationStatus`	Darwin Core	reused as the automation boundary
`dataGeneralizations`, `informationWithheld`	Darwin Core	reused for coordinate semantics and sensitive taxa
`samplingProtocol`, `samplingEffort`, `sampleSizeValue`/`Unit`	Darwin Core + Humboldt	reused; effort reported per modality, never summed
Inventory completeness, sampling scope	Humboldt	partially used
Device model, sensor specification	SensorML concepts	re-expressed in platform tables (not SensorML-encoded)
Derivation chain (raw → detection → identification)	PROV-DM concepts	re-expressed in remarks fields (not PROV-encoded)
Platform, mission, annotation documentation	FAIR² Drones	dataset card template adopted
Release principles	FAIR, CARE, TRUST	followed
Duty cycle and acquisition schedule	—	systems extension (deployment `samplingProtocol` + config tables)
Trigger semantics (PIR sensitivity, burst length, delay)	—	systems extension (config tables, S2.3)
Timestamp confidence and repair method	—	systems extension (timestamp map, S5)
Per-detection coordinate provenance status	—	systems extension (`drone_georef_status`, S6)
Typed missingness	—	systems extension (S8)
Model identity, version, parameters, per-rank accuracy	—	systems extension (S9)
Validator coverage and known gaps	—	systems extension (S11)

The last six rows are the ones no standard we found expresses in a machine-readable way. We name them and use them consistently; we do not claim to have standardized them.

S8. Missing data, broken down by type.¶

#	State	How it is computable	Example
1	Sensor active, nothing observed	Hourly bin inside a deployment window with no Occurrence	Most nocturnal camera bins
2	Sensor inactive by schedule	Deployment-level schedule field	5 min/hr recorders are off 91.7% of the time by design
3	Sensor inactive by failure	Capture gap vs. logged deployment window, with cause	TW02-CT01 battery, Aug 18–25
4	Captured but not represented	Raw file inventory vs. Event table	78 transfer-duplicate WAVs and 4 failed WAVs, each classified per path (S3)
5	Metadata unusable for a purpose	Status/provenance column	70,219 drone detections unprojectable (S6); until 2026-11-10, 41,524 drone Occurrences with a date but no time of day (S10, since recovered)
6	Excluded in QC	Documented exclusion	People-only TW07-CT05 deployment excluded; one drone video reviewed and kept as a zero-Occurrence `MediaEvent`

States 1 and 2 have opposite implications for a coordination policy, which is why the schedule field is load-bearing and not documentation.

S9. Models, parameters, and measured accuracy¶

Manual annotation does not scale to sensor-network volume: human-reviewed labels are under 1% of the release and consumed the review capacity available. Semi-automated analysis is therefore the realistic condition, and we use current state-of-the-art models so that reported quality reflects what a practitioner adopting today's best tooling would obtain.

Labels were produced in two generations, and the released tables carry both. The first generation labeled every SAM3 box with a BioCLIP Tree-of-Life taxon; it still describes the operational drone table and the merged cross-modal snapshot. The second generation (October–November 2026) replaced the camera-trap labels with embedding-based triage plus targeted human review (S9.1), and added a seven-video human-annotated drone reference set with a manual starling supplement (S9.2).

Stage	Model	Role
Visual segmentation	SAM 3	bounding boxes on camera-trap images and drone frames
Visual classification, first generation	BioCLIP Tree-of-Life (TreeOfLife-200M)	top-1 taxon with full rank hierarchy, top-5 retained; operational drone labels
Label triage, second generation	Resolve–Clear–Defer prototype triage over proposal embeddings	routes proposals to human review and resolves the rest by similarity to reviewed examples; operational camera-trap labels
Acoustic detection	BirdNET	detections with frequency bounds and offset into the WAV
Water/lake masking	SAM 3	lake/water masks for habitat-association analysis

S9.1 Camera-trap labels (post-triage)¶

Operational camera-trap Occurrences are SAM3 animal proposals labeled by Resolve–Clear–Defer triage. The triage selects proposals for human review and resolves the remainder by similarity to reviewed examples; labelProvenance records the outcome on every row, with the triage round where recovered.

`labelProvenance`	`identificationVerificationStatus`	Occurrences
`triage_reviewed`	`human-reviewed`	1,760
`triage_resolved`	`machine-resolved`	70,808
`triage_forced_1nn`	`machine-resolved`	201
Total		72,769

Forced nearest-neighbour labels did not meet the triage confidence rules and may be wrong; exclude them unless label noise is tolerable. Box geometry is never edited (geometryProvenance = SAM3), even when the label was human reviewed. The table keeps only reviewed or resolved animal proposals: cleared and background proposals, boxes with a shorter side below 32 px, boxes overlapping the per-resolution ROI mask, and frames on the privacy exclusion manifest are excluded, so an image without an Occurrence is not evidence that no animal was present. Twelve taxa occur: milu (Père David's) deer, Przewalski's horse, Canada goose, Sichuan takin, trumpeter swan, killdeer, European starling, barn swallow, and the broad labels bird, butterfly, dragonfly, and insect.

A separate human-reviewed reference set of 2,067 boxes on 198 selected ungulate and waterfowl images (881 Canada goose, 627 milu deer, 349 Przewalski's horse, 149 Sichuan takin, 61 trumpeter swan) is released beside, not inside, the Occurrence table. Taxonomy and geometry were both reviewed; 1,552 boxes share a SAM3 proposal's geometry and 515 are human edits or additions, and the set is not filtered by the 32 px or ROI rules. The images were selected for review, so the set is not a random or exhaustive test split. Triage label accuracy against this reference set has not yet been reported.

S9.2 Drone labels: operational table and seven-video reference set¶

Operational table. The 318,203 machine drone Occurrences remain first-generation labels: a SAM3 box on every 15th frame with a top-1 BioCLIP Tree-of-Life taxon. The only measured accuracy for that labeler comes from human-annotated camera-trap subsets:

Analysis	Level	n	Precision	Recall	F1 / accuracy
SAM3 animal localization, camera traps	IoU ≥ 0.50	1,125	0.866	0.845	0.856
Tree-of-Life taxonomic assignment	class	951			0.845
	order	951			0.660
	family	951			0.463
	genus	951			0.322
	species	951			0.271

Usable statement: class-rank labels support screening and triage; species-rank labels require curation; anything aggregating across ranks is uninterpretable. A single headline accuracy anywhere in the 0.27–0.85 range would mislead. Corroborating evidence: the taxonomy crosswalk holds 12,577 label keys, of which only 4 consist entirely of expert-verified records and 16 contain accepted names with mixed verification evidence.

Seven-video reference set. Seven drone videos were human-annotated on 1,089 sampled frames for four focal classes, with onager observations excluded: 20250703_parrotanafi_1, _2, _5 (C1), 20250811_DJI_01, 20250818_parrot_01, _02, and 20250825_parrot_01 (C2). The corrected reference holds 27,159 four-class boxes (the localization audit below was run against 27,180 before the final manual review): Père David's deer 21,177, Sichuan takin 3,293, Przewalski's horse 2,066, Canada goose 644. It is delivered as COCO JSON with provenance, a frame-to-proposal mapping, and the audit files.

SAM3 localization against the reference (35,312 proposals, one-to-one matching):

IoU	Precision	Recall	F1
0.25	0.758	0.985	0.856
0.50	0.749	0.973	0.847
0.75	0.740	0.961	0.836

Per-class recall at IoU 0.25: Przewalski's horse 0.961, Père David's deer 0.985, Sichuan takin 0.993, Canada goose 1.000. Per-video recall at IoU 0.25 ranges from 0.958 (20250818_parrot_02) to 0.999 (20250825_parrot_01).

Resolve–Clear–Defer prototype triage was then run over the full proposal pool (seed 3, density-moderated k-center acquisition, calibrated rule k = 5, τ_fg = 4, τ_bg = 4). It spent 100 proposals on calibration and 1,610 on expert review (4.6% of the 35,212-proposal evaluation pool), handled the remaining 33,602 automatically, and got 28,040 of those right: 79.6% of the evaluation pool, or 83.5% of the automatic decisions. Zero proposals were left deferred.

Four-species classification on the 25,897 matched, unreviewed foreground instances (78.4% Père David's deer, so always-deer scores 0.784 accuracy):

Method	Precision	Recall	F1	Accuracy	Macro-F1
Prototype triage, automatic decisions only	0.913	0.850	0.880	0.850	0.789
BioCLIP 2 zero-shot, four prompts	0.167	0.167	0.167	0.167	0.230
Tree-of-Life, open-world, exact focal-taxon match	0.947	0.037	0.071	0.037	0.188

Species-wise recall for triage: Père David's deer 0.918, Przewalski's horse 0.687, Canada goose 0.590, Sichuan takin 0.551. The dominant automatic errors are wrong-species resolves into Père David's deer (takin 1,287, horse 595, goose 100) and 1,584 deer cleared as background. Read the table carefully: Tree-of-Life is evaluated open-world, so its 0.037 recall is exact top-1 recognition of a focal taxon under an unconstrained vocabulary, not animal recognition accuracy; BioCLIP 2 zero-shot always answers one of four prompts and has no background option; triage abstentions count as no prediction. All figures are proposal-level repeated observations from seven videos, not independent animals.

Manual European starling supplement. Twelve human-added Sturnus vulgaris boxes on one frame (20250703_parrotanafi_2 / VID_002, frame 1320) were appended to the drone Occurrence table on 2026-11-01 with basisOfRecord = HumanObservation, identificationVerificationStatus = human-supported; limited-resolution, and remarks stating that no SAM3 proposal, embedding, or triage decision was involved. The identification rests on limited pixels in a single frame and should be interpreted cautiously. These records are excluded from the 27,159-box reference, the localization audit, the triage calibration and evaluation, and the species-classification comparison; none of the numbers above change if they are dropped. The 12 rows were appended twice (24 rows, 12 distinct occurrenceIDs) in both the drone and merged tables (defect 21, repaired in the release build).

S9.3 Verification status distribution¶

Counts are per table, because the vocabularies differ. The merged table below is the pre-update state (2026-11-01), before the camera rows were regenerated from the post-triage tables (defect 22, resolved 2026-11-08).

Merged occurrences_all_modalities.csv as of 2026-11-01:

Modality	Status	Records
Camera traps (first generation)	`unverified`	223,203
	`validated`	946
Bioacoustics	`unverified`	9,960
Drones	`unverified`	317,559
	`validated`	644
	`human-supported; limited-resolution`	24 (12 distinct records; defect 21)
Citizen science	`verified by expert`	622
Text data	`verified by expert`	30
Total		552,988

Post-triage camera-trap Occurrence table (the released camera dataset):

Status	Records	Share
`machine-resolved`	71,009	97.6%
`human-reviewed`	1,760	2.4%

validated = 946 camera-trap + 644 drone bird records from the curated goose/swan review set (added 2026-09-20). The 946 camera records belong to the first-generation camera table and are not carried into the post-triage Occurrence table; the 2,067-box reference set takes their role. verified by expert = all 622 citizen-science + 30 text-data records.

A verification floor is one predicate: treat unverified and machine-resolved as machine labels; human-reviewed, validated, and verified by expert as person-reviewed; and human-supported; limited-resolution as person-reviewed but low confidence.

S9.4 Why this must be published¶

A taxon string is not reproducible without the model that produced it. Two versions of the same classifier yield different taxa from the same image.
Published uncertainty is what makes labels re-derivable. Machine labels are a snapshot of a model, not ground truth. A release that keeps source media, detection geometry, model identity, and per-rank accuracy can be re-labeled when better models arrive and the label sets compared meaningfully. One shipping only final taxon strings cannot.
Aggregate accuracy hides the rank-dependent failure mode above.
Verification status must be per record, not per dataset — a README caveat cannot be filtered on.

S10. Sentinel values: `nan` as a date string (repaired)¶

In snapshot darwin_core_latest_20260724, the 2025-08-11 drone session carried eventDate = the literal string nan (a dataframe export of a missing value) and an empty eventTime. The QC summary reported missing_eventDate=0 because the check tested for null/empty and "nan" is neither.

eventID	Occurrences
`TheWilds_2025_drone_20250811_DJI_01_DJI_0002`	12,522
`TheWilds_2025_drone_20250811_parrot_01_P5190525`	12,430
`TheWilds_2025_drone_20250811_parrot_01_P5190526`	11,996
`TheWilds_2025_drone_20250811_DJI_01_DJI_0001`	4,535
`TheWilds_2025_drone_20250811_parrot_01_P5180524`	41
`TheWilds_2025_drone_20250811_DJI_01`, `..._parrot_01` (`SensorDeployment`)	—
Total	41,524 (13.1% of drone Occurrences)

Repair (release build 2026-10-22). eventDate was restored to 2025-08-11 from the flight identifier and marked as identifier-derived. Drones flew on 6 dates; the old count of 7 included the string nan.

Time recovery (2026-11-10). The processed telemetry had no clock, but the five video containers carry a creation time. Whether an aircraft writes UTC or local time into that field is airframe-specific, so it was settled against flights with independently known times: the Parrot Anafi writes UTC (its 2025-08-18 video agrees with the telemetry file name to within 2 s), and the DJI Air 2 writes UTC (its 2025-09-08 video agrees with the telemetry's datetime(utc)). The 2025-08-11 Parrot videos therefore started at 09:12:43, 09:14:19, and 09:19:55 local. The DJI Mavic 3 videos read 10:29:39 and 10:31:32; read as UTC that would be 06:29, before anyone was on site, so this airframe wrote local time, which the telemetry export file name (…_[10-29-13]) and the field log also support. The field log places the Parrot flight at Stop 1 (Nomad Ridge, lake survey) and the DJI at the dock, Stop 5; the JotForm check-ins run 09:32 to 10:48. eventTime is now set for the 5 MediaEvents and 2 flights (start/end intervals) and for all 41,524 Occurrences (video start + frame_index / fps, to the second). The DJI convention is inferred rather than verified, so those times should be read as minute-level. The record is metadata/provenance/drone_time_recovery_20250811.json in the drone dataset.

The same audit found impossible coordinates (nan, longitude ≈ 500) on seven drone Events, now replaced with a validated static video location or left blank, and a missing study-root Event TheWilds_2025, now added.

S11. Validation: checks run, and gaps¶

S11.1 Original checks (snapshot `20260724`)¶

no missing event links (eventID resolves)
no missing taxon
no missing eventDate (presence only; this is how the nan dates passed)
no duplicate occurrenceID

S11.2 Current release checks¶

The generator now parses every date, time, and coordinate pair and rejects sentinel values (nan, None, NULL, -1, 1970-01-01) before export. The release audit checks, over every modality and merged table:

source-to-merged parity for Events and Occurrences;
a single-rooted Event hierarchy with no dangling parentEventID;
Occurrence-to-Event links and ancestry;
parsed dates and times, and WGS84-bounded coordinates;
every Occurrence date inside its parent Event's span and the study window;
no absolute cluster paths, and every associatedMedia resolving to a released file in the dataset's media manifest;
agreement between the generator and the released tables.

S11.3 Gaps that remain¶

Gap	Consequence
No check that analyses constrain `drone_georef_status`	Mixed-semantics spatial figures are possible (S6)
No semantic content check	People in imagery are found by a person detector plus human review, not by the validators

A passing validator is evidence about the validator. The original checks certified 552,964 records clean while 13.1% of one modality had no usable timestamp. Validator coverage should be reported, not assumed.

S12. Darwin Core field dictionary¶

occurrences_all_modalities.csv — 44 columns:

modality_type, occurrenceID, eventID, catalogNumber, scientificName, kingdom, phylum, class, order, family, genus, specificEpithet, taxonRank, vernacularName, occurrenceStatus, individualCount, lifeStage, sex, behavior, basisOfRecord, recordedBy, identifiedBy, identificationVerificationStatus, identificationRemarks, eventDate, eventTime, decimalLatitude, decimalLongitude, coordinateUncertaintyInMeters, geodeticDatum, locality, locationID, country, stateProvince, county, establishmentMeans, degreeOfEstablishment, occurrenceRemarks, associatedMedia, associatedReferences, datasetID, datasetName, informationWithheld, dataGeneralizations

events_all_modalities.csv — 21 columns:

eventID, parentEventID, eventType, eventDate, samplingProtocol, sampleSizeValue, sampleSizeUnit, locality, locationID, decimalLatitude, decimalLongitude, coordinateUncertaintyInMeters, geodeticDatum, country, stateProvince, county, datasetID, datasetName, recordedBy, fieldNotes, eventTime

S12.1 Event hierarchy composition¶

`eventType`	Rows	Granularity
`Project`	6	study root `TheWilds_2025` + one per modality
`SensorDeployment`	37	12 camera + 10 ARU + 15 drone flight session
`AudioRecording`	2,371	one per processed WAV file (64 with no detections)
`MediaEvent`	39	one per raw drone video (1 reviewed, no Occurrences)
`ObservationSession`	17	13 citizen science + 4 text data
Total	2,470

S12.2 Occurrence composition¶

Modality	Events	Occurrences	Dates	Hourly bins	`locationID`s
Study root	1	—
Camera traps	13	224,149	29	624	7
Bioacoustics	2,382	9,960	29	650	6
Drones	55	318,203	6	10*	15
Citizen science	14	622	13	28	0
Text data	5	30	4	4	6
Total	2,470	552,964

* includes the 2025-08-11 session's bins after its times were recovered on 2026-11-10 (S10); the integration pass had counted 8. Camera dates and bins use the corrected timestamps of the release build (439 curated records re-timed).

Camera-trap figures here are the first-generation table in the merged snapshot. The released post-triage camera dataset has 19,762 Events (1 Project, 12 SensorDeployment, 19,749 Image) and 72,769 Occurrences, and the drone table gained 12 manual starling records on 2026-11-01 (S9, defects 21–22).

S13. Arrival characterization, both layers¶

Window: Campaign 2, 2025-08-04 to 2025-08-28, 600 hours — the only interval in which all four modalities were simultaneously active.

Definitions. Acquisition arrival = one image capture (from camera_image_timestamp_map.csv, corrected timestamps) or one WAV recording (from AudioRecording Event times). Detection arrival = one Occurrence with a resolvable date and time. B = (σ−μ)/(σ+μ) = (CV−1)/(CV+1) over inter-arrival times: −1 periodic, 0 Poisson, →1 maximally bursty.

Stream	Arrivals	Rate/hr	CV	B	Dates
Acquisition
Camera image captures	220,697	367.83	2.06	+0.347	25
Acoustic WAV recordings	2,060	3.43	1.23	+0.103	25
Detection
Camera traps	208,054	346.76	7.26	+0.758	25
Bioacoustics	8,865	14.78	3.21	+0.525	25
Drones	94,762	157.94	216.25	+0.991	3
Citizen science	242	0.40	5.30	+0.682	10

Detection streams count machine-derived Occurrences only; curated review records are excluded. Acoustic recordings include the 64 zero-detection WAVs. The drone row includes the 41,524 Occurrences of 2025-08-11 since their times were recovered (S10); before that recovery the stream had 53,238 Campaign-2 arrivals at 88.73/hr, CV 229.61, B +0.991 over 2 dates. Every stream is released as a versioned trace (analysis/arrival_traces_v2/ in the study-level dataset) with a source-record ID per arrival, inclusion rules, and input checksums; earlier unversioned arrival summaries are deprecated. Acquisition arrivals are not reconstructible for drones (per-frame times are not retained) or citizen science (one submission per session).

S13.1 Diurnal profiles (mean arrivals per hour-of-day over 25 days)¶

Hour	Cam capture	Cam detect	Acoustic WAV	Acoustic detect	CS detect
00	290.1	76.5	1.8	3.4	0.0
01	296.3	86.9	1.8	3.2	0.0
02	317.4	184.1	1.8	3.3	0.0
03	300.3	128.6	1.8	4.3	0.0
04	304.0	211.9	1.8	3.3	0.0
05	300.6	133.3	2.8	21.1	0.7
06	317.9	519.2	5.2	41.6	0.0
07	321.8	674.4	4.3	20.1	0.0
08	341.3	503.2	4.3	18.3	0.0
09	407.2	597.6	4.6	18.4	1.7
10	454.4	416.5	4.3	18.6	2.7
11	534.4	495.1	4.3	16.8	1.4
12	589.0	507.6	4.3	16.2	0.2
13	611.8	496.8	4.3	16.2	0.4
14	527.6	301.1	4.3	15.8	0.0
15	446.0	285.0	4.3	14.6	0.0
16	385.3	193.5	4.3	16.5	0.0
17	326.1	536.7	4.3	16.4	0.0
18	297.8	499.9	4.3	17.6	0.0
19	286.4	393.7	5.2	36.2	0.0
20	292.8	494.8	3.0	18.7	0.0
21	304.8	392.9	1.8	7.2	0.0
22	288.4	138.9	1.8	3.0	1.6
23	286.1	53.9	1.8	3.8	1.0

Drone detections are zero in every hour but 09 (26,619 arrivals summed across the two flight dates).

Diurnal swing, max/min: camera captures 2.1×, camera detections 12.5×, acoustic acquisition 2.9×, acoustic detections 13.9×. Detection profiles are sharper than the acquisition profiles feeding them in every modality where both are measurable.

These values come from the versioned v2 arrival traces. They supersede the earlier unversioned arrival summaries, whose event definitions were not recorded.

S13.2 Field servicing effort¶

Derived from the Campaign-2 Jotform logs in thewilds_20250804_20250828/jotforms/ (two camera-trap exports, two bioacoustic exports). Every row is one sensor servicing action with a DateTime stamp, site ID, device ID, SD-card ID, team member, and free-text note.

Trip	Actions	Window	Elapsed	Median gap between actions	People
2025-08-04	8	08:58–10:41	103 min	16.0 min	3
2025-08-07	6	08:59–10:05	66 min	3.0 min	1
2025-08-11	9	09:32–10:48	76 min	3.0 min	2
Total	23		245 min (4.1 h)

Breakdown of the 23 actions: 11 dropoff (deployment), 6 pickup (retrieval), 5 check-in (status only, no SD or battery swap), 1 combined. By modality: 12 acoustic, 11 camera.

Additional operational fields captured per action:

Access difficulty, 1–3 scale — recorded on 15 of 23 actions (rating 3: 9 actions; rating 1: 5; rating 2: 1)
Battery status — e.g. 98%, 95%, 60%, "Low", "100%"
SD-card storage — e.g. "7% full", "26% full", "5GB"
Camera heading — compass octant, used by the camera-trap georeferencing
Relocation notes — free text landmarks ("Zebra shelter", "Zip tower", "Lake trail"), which is how the TW09-CT02 mid-deployment move was recovered
Team member name and email

Scope boundary: this is measured elapsed field time for the three retained mid-campaign Jotform exports only. Campaign-1 field notes have no clock times, Aug. 28 retrieval submissions are absent from the retained exports, and no per-stage analyst-time log exists. The release therefore reports no total campaign field effort and no measured desk-processing hours.

Coverage limits. These exports cover the three mid-campaign trips only. Campaign-1 field notes (thewilds_20250630_20250703/field_notes/*.md) are narrative and contain no clock times, and the 2025-08-28 retrieval visits do not appear in the exported forms. So 4.1 h is logged servicing time, not total field effort.

Not recorded anywhere: analyst time per processing stage. Field visits were logged with timestamps; the timestamp-repair and taxonomy-review work was not. This asymmetry is requirement R7 in the paper.

S14. Reproduction¶

Darwin Core generation:

The scripts live in the SmartWilds code repository (scripts/). Set SMARTWILDS_ARCHIVE to the root of the field archive before running them; see scripts/README.md.

python scripts/darwin_core/generate_darwin_core.py --force
python scripts/arrival_traces/generate_arrival_traces_v2.py

Camera-trap georeferencing:

python scripts/georeferencing/geolocate_camera_trap_treeoflife_predictions.py \
  --input-dir  <archive>/processing_output/detections/camera_traps/methods_paper/annotations/sam3/treeoflife_sam3/predictions \
  --config-csv scripts/georeferencing/camera_geolocation_config.csv \
  --output-dir <archive>/processing_output/detections/camera_traps/methods_paper/annotations/sam3/treeoflife_sam3/geolocation/geolocated_predictions \
  --overwrite

Drone georeferencing:

python scripts/georeferencing/geolocate_drone_treeoflife_predictions.py \
  --prediction-dir <SAM3 per-video prediction CSVs> \
  --detection-dir  <archive>/processing_output/detections/drones \
  --config-csv     scripts/georeferencing/drone_projection_config.csv \
  --output-dir     <archive>/processing_output/detections/drones/geolocation/geolocated_predictions \
  --overwrite

S14.1 Release discipline¶

Integration passes are dated and immutable. Every mutation retains a before-copy and emits a JSON audit record. Passes differ semantically, not just additively — different visual-label semantics, corrected timestamps — so comparing numbers across passes is invalid in a way that is not visible in the numbers. Any figure or table must name the release build it was computed from.

The 2026-10-22 release build supersedes the 2026-07-24 pass; each of its repairs (S10, S3, S4, S16) has its own JSON audit record, shipped under metadata/provenance/. The 2026-09-20 curated-bird update is reconstructible from:

curated_bird_dwc_update_audit_20260920_015706.json
curated_bird_event_sync_audit_20260920_020046.json
curated_bird_project_sample_size_fix_20260920_020401.json

S15. Release logistics¶

Each modality is published as its own Hugging Face dataset with its own DOI, and a study-level dataset holds the merged Event table, the text records, the arrival traces, and this supplement. The SmartWilds collection joins all five. DOIs, the license, and the maintenance commitment will be listed here at publication. Contact: Jenna M. Kline (kline.377@osu.edu).

S16. Defect inventory, in full¶

Every defect class found while integrating the SmartWilds campaigns (release build 2026-10-22; camera tables updated 2026-11-08). Detected by is what made the defect visible, grouped into five families: an external record (service log, field note, platform documentation), a redundant encoding inside the data, a cross-check between two artifacts, human inspection, or nothing. Rows 19-20 had no second record and could be bounded but not repaired. Rows 21-22 were found after the 2026-11-01 annotation update and were repaired in the release build (row 21) and the 2026-11-08 camera update (row 22).

#	Defect	Mod.	Scope	Detected by (family)	Repair	Residual risk
1	Device clock wrong or unset	Cam	221,306 / 235,778 images (93.9%)	Service log deploy/retrieve times (External record)	18 per-deployment reconstruction methods (paper §timestamp reconstruction; S5)	130,290 images (55.3%) only medium confidence
2	Clock reset to factory epoch	Cam	2 deployments, 1 reporting year 2023	Service log; implausible date on its face (External record)	Anchored offset from known deploy time	Offset assumed constant over deployment
3	Media split across SD cards with inconsistent clocks	Cam	TW02-CT01 (4 clock states), TW09-CT02 (2), TW06-CT04 (2), TW04-CT03 (2)	Card-level directory structure + service log (External record)	Per-card anchors reconciled against logs	Within-card drift unmodeled
4	Timestamp only in burned-in overlay	Cam	182 images (TW06-CT05, TW09-CT05, TW11-CT06)	The image itself (redundant in-frame encoding) (Redundant encoding in data)	OCR of bottom overlay; 3 read manually	OCR errors on 179 automated reads
5	Camera mounted inverted	Cam	TW04-CT04, 570 images	Human review of imagery (Human inspection)	Flagged; images rotated downstream	---
6	Camera relocated mid-deployment	Cam	TW09-CT02 (dock → Nomad Ridge W, Aug 18)	Field note recording the move (External record)	Split into two logical location intervals	Single `locationID` spans both
7	Battery exhaustion	Cam	TW02-CT01 Aug 18--25; TW02-CT03 C1 (55 files)	Capture gap vs. logged deployment window (Cross-check)	Recorded as a failure gap with cause	Lost observation window unrecoverable
8	Non-wildlife content (visitors)	Cam	TW07-CT05, whole deployment	Human review of imagery (Human inspection)	People-only imagery not uploaded; deployment excluded from release	Card disposition after field not documented
9	Recorder ran a schedule other than configured	ARU	TW05-SM01 C1 (144 files)	File inventory vs. intended config (Cross-check)	Actual schedule inferred and documented	Intended vs. actual effort differ
10	WAV files with no event record	ARU	146 / 2,453 paths (6.0%)	File inventory vs. event table (Cross-check)	Classified per path: 64 zero-detection recordings added as Events; 78 transfer duplicates and 4 failed files excluded	4 fragments/unreadable files unrecoverable
11	Heterogeneous telemetry schemas	Drone	3 platforms, 3 formats, feet vs. metres	Platform documentation (External record)	Per-platform field mapping with unit conversion	Mapping assumptions per platform
12	Drone video not processed	Drone	1 / 39 raw videos	File inventory vs. event table (Cross-check)	Reviewed (no usable content); retained as a zero-Occurrence `MediaEvent`	---
13	Media references to purged scratch workspace	Drone	16,069 frame paths, 307,260 Occ.	Resolving every `associatedMedia` path (Cross-check)	Replaced with the retained original video, flagged as a video-level fallback	Frame-level media link lost for those rows
14	Unreviewed fine-rank taxon labels	Cam, Drone	Merged snapshot: 550,722 Occ. `unverified` (99.6%); post-triage camera table: 71,009 Occ. `machine-resolved` (97.6%)	Human-annotated validation subsets (Human inspection)	Camera labels replaced by embedding-based triage with human review and a 2,067-box reference set (S9.1); drone labels kept `unverified`, with a seven-video human-annotated reference set (S9.2); accuracy reported per rank and per method	Drone species-rank labels include implausible taxa; triage label accuracy against the camera reference set not yet reported
15	Null written as the string `nan`	Drone	41,524 / 318,203 Occ. (13.1%); 5 Events + 2 deployments	Date redundantly encoded in the `eventID`; time from the video containers cross-checked against the field log (Redundant encoding in data; External record)	Date restored from the identifier and flagged as derived (release build 2026-10-22); time of day recovered on 2026-11-10 from the video container creation times, with each aircraft's clock convention settled against reference flights and the field log and JotForm check-ins of the day (`provenance/drone_time_recovery_20250811.json`)	DJI clock convention is inferred from the field schedule, not verified on a reference flight; treat those times as minute-level
16	Impossible drone coordinates (`nan`, longitude ≈500)	Drone	7 Events	Domain check on coordinate pairs (Cross-check)	Valid static video location if inside study area, else blank	Those Events lack a flight-level position
17	Curated records bypassed timestamp correction	Cam	439 / 946 curated Occ.	Occurrence date outside parent deployment interval (Cross-check)	Re-timed from the corrected camera timestamp map (release build 2026-10-22)	---
18	Root Event row absent from Event table	All	1 dangling `parentEventID`	Referential check over the Event table (Cross-check)	Study root Event added; parent links validated	---
19	Within-card clock drift	Cam	Up to 130,290 images	none (None (recorded once))	Constant clock rate assumed	Unquantifiable timestamp error
20	True deployment / retrieval times	Cam	Deployment metadata known to contain errors	none beyond the erroneous records (None (recorded once))	Corrected occurrence dates used instead	True camera-days not reportable
21	Manual starling supplement appended twice	Drone	24 rows / 12 distinct `occurrenceID`s in the drone and merged Occurrence tables (2026-11-01)	Duplicate `occurrenceID` check (Cross-check)	Repaired in the release build: the 12 exact duplicate rows are dropped, leaving every `occurrenceID` unique (318,215 drone Occurrences)	---
22	Merged snapshot and camera release disagree on camera labels	Cam	224,149 first-generation vs 72,769 post-triage Occ.; 13 vs 19,762 camera Events	Row-count parity between the camera release tables and the merged table (Cross-check)	Not yet repaired; regenerate the merged snapshot and the camera detection-layer arrival traces from the post-triage tables	Cross-modal totals (552,964) and camera detection-layer traces describe the first-generation labels