Use edge processing for timely cues and central processing for deeper checks, retaining the original evidence.
Keep time-sensitive cues near the sensor
Onboard detection can provide timely candidate and quality alerts in poor connectivity. Research from Namibia illustrates lightweight inference for conservation in such settings. A detection must not automatically authorize a closer approach. Keep welfare and flight constraints authoritative, and prevent perception failures from propagating into flight-critical control.
Measure latency, power and misses together
Compression, quantization and reduced input resolution may improve throughput while losing small targets. Measure latency, dropped frames, temperature, power, endurance and class-specific errors on deployed hardware. Desktop speed is not onboard performance. Record skipped inference periods so a lack of alerts is not interpreted as an absence of animals.
Centralize deeper processing and review
Central processing can support higher-resolution inference, sensor fusion, cross-sortie reconciliation, model comparison and collaborative review. Use resumable transfers and integrity checks with an upload ledger. Decide which sensitive information may be transferred and where it may be stored before uploading. Connectivity does not remove governance requirements.
Retain enough evidence to audit negatives
Transmitting only detections saves bandwidth but can make false negatives unauditable. Define retention of originals, negative samples and decision logs beforehand. Compare transmission savings with future review, retraining and reacquisition costs. Aircraft, field workstation and cloud may work better together than an all-edge or all-cloud approach.
Evidence & further reading
Check the original study conditions, species and location before using a result in a plan. Do not transfer findings universally across species and settings.