السياق
Saiga move across steppe that is too large for a reliable ground count. A detector that marks animals in video does not define the number of saigas. That number is a measurand: the population, the animals available to the sensor, the sample, the detections, the unique tracks, and a design-based estimate are different quantities.
حدود النظام
ضمن النطاق
- optical and thermal capture
- ground sampling distance and feature-pixel limits
- detection and instance segmentation
- a male–female–unknown sex gate
- tracking and deduplication
- ray–terrain geolocation
- conservation records for the three Kazakhstan population systems
خارج النطاق
- flight control
- airframe design
- enforcement
- any abundance or sex-ratio claim that has not passed a locked field test
البنية
The chain follows the September 2026 manuscript. Biological identity comes before coat colour. Camera height, sensor width, focal length, and blur decide whether a feature is observable. A versioned detector and an independent segmentation baseline emit observations. Tracking removes repeated views of the same animal. A calibrated camera ray and a terrain model place the observation; a versioned GIS layer assigns population system and administrative unit as separate attributes. Survey design, not model confidence, is what could expand a visible count.
- Biology
- Observability
- Detect / segment
- Track
- Geolocate
- Records
- Train ≠ serve
- Detections are events
- Biology
- Observability
- Detect / segment
- Track
- Geolocate
- Records
أعداد من السجل المنشور. تصف البنية، وليست نتيجة مقاسة.
أين يصبح تشغيل النموذج حدثاً؟
field
- خارجيBiology
- واجهةObservability
inference
- عمليةDetect / segment
- عمليةTrack
ground
- عمليةGeolocate
handoff
- مخزنRecords
- BiologyObservability
- ObservabilityDetect / segment
- Detect / segmentTrack
- TrackGeolocate
- GeolocateRecords
التدفقات الحرجة
A planned pass records pose, exposure, and frames
geometry decides what is observable
models detect and segment
sex stays unknown unless the head is resolved
a lost track is closed rather than merged
coordinates and uncertainty are written
conservation records consume unique observations.
المراحل
القرارات
Shape before colour
Coat colour changes with season, dust, snow, and exposure. The nasal profile and body form are the species evidence. Horns support sex only when the head is resolved.
Unknown is a measurement
Sex is male, female, or unknown. Forcing a difficult animal into the nearest class biases the demographic record.
A box is not a census
Detection, unique identity, and abundance are separate stages. Model confidence is not an inclusion probability.
Keep flight out of inference
The model must not own the aircraft. Capture is an input; control stays a separate system.
عندما يتعطل
| الحالة | الكشف | الاستجابة |
|---|---|---|
| Feature pixels fall below the task | Ground sampling distance against the prevalidated minimum | Refuse head or horn claims; keep the animal unknown |
| Empty or stalled capture | Frame-rate and stream health | Record a coverage gap; do not invent detections |
| Thermal and optical desync | Timestamp window between bands | Hold the lagging band; mark track confidence |
| Lost track across a pass | Track identity timeout | Close the track and start a new identity rather than merge two animals |
| Coordinate uncertainty crosses a boundary | Error ellipse against the polygon edge | Flag the assignment; do not treat extra decimal places as accuracy |
البنية التحتية
- Field UAV
- calibrated optical and optional thermal payloads
- inference runtime
- prediction ledger
- versioned GIS
- conservation analytics surface
النتيجة
Conservation teams receive detections, tracks, and georeferenced observations whose limits are explicit. Analytical geometry can be checked. Field accuracy, sex coverage, and abundance stay unclaimed until a locked test exists.
الحالة الحالية
The monitoring path follows the September 2026 methodological manuscript. Analytical geometry is checked. Field accuracy, sex coverage, and abundance remain unmeasured.