AVIX-AI 4-Stage Pipeline: How Vertiports Kill Bird Risk Before Wheels-Up

Abstract

Bird strikes are not an airport legacy problem that shrinks when aircraft get smaller. They scale in the opposite direction: eVTOL rotors spin at higher RPM than turbofan fan stages, ingest smaller debris volumes than a CFM56, and carry no Bird Ingestion certification margin equivalent to FAA AC 33.76. For a Korean vertiport operator approaching the K-UAM 2027 commercial window, that asymmetry is existential. The question is not whether wildlife hazard management is required — ICAO Doc 9332 and Korea Airports Corporation (KAC) operational standards already mandate it — but whether the documentation chain is machine-auditable enough to survive a post-incident regulatory review or a pre-certification site inspection. UAM KoreaTech's AVIX-AI BirdThreat pipeline answers that question with a four-stage operating doctrine: Identify · Classify · Respond · Monitor. Validated at 19/19 HTTP 200 across all test cycles at Incheon Technopark (commit fbcb327, 2026-04-20), the pipeline runs on a Jetson Orin Nano edge node, publishes animal-class entities natively into Anduril Lattice, and produces a timestamped audit trail that manual wildlife logs structurally cannot replicate. This article decomposes each stage for vertiport ground-operations teams and K-UAM working-group members scoping the 2027 certification runway.

---

1. Operational Anchor — Incheon Technopark Validation Run

The Site

Incheon Technopark sits within the Songdo International Business District, approximately 8 km from Incheon International Airport's southern perimeter. The campus hosts mixed-use rooftop and ground-level open space that mirrors the vertiport typology — a hard-surface landing pad surrounded by landscaped habitat margin — making it a structurally valid test environment for ground-level wildlife detection. Incheon is also administratively within the Incheon Free Economic Zone (IFEZ), which grants accelerated permit processing for aviation-adjacent technology trials. The combination of proximity to a major international airport's wildlife pressure zone and IFEZ regulatory flexibility made Incheon Technopark the highest-signal site available for a pre-commercial validation run.

Environmental Read

The Incheon coastal corridor sits directly on the East Asian–Australasian Flyway (EAAF), one of nine major global migratory flyways tracked by the EAAF Partnership. During spring and autumn migration windows, shorebird and waterfowl concentrations at tidal flats within 15 km of the Technopark site routinely exceed 10,000 individuals per survey count. Resident corvid and pigeon populations maintain year-round pressure. This is not background noise — it is the baseline environmental load against which any vertiport wildlife management system must demonstrate reliable discrimination and response. A system that performs at 90% detection in a low-pressure environment and degrades in peak-flyway conditions provides no certification value.

Differential Factor

What distinguishes the Incheon Technopark run from a laboratory benchmark is the operational fidelity of the failure mode. The 19/19 HTTP 200 result was not achieved under controlled low-clutter conditions. The site introduces ambient RF interference from adjacent industrial units, variable coastal wind loading that degrades acoustic sensor baseline, and transient human movement that challenges entity classification. A zero-silent-failure result in that environment is not a marketing claim — it is the minimum bar required to assert that the pipeline does not produce invisible gaps in the audit record under real Korean coastal conditions.

Modern Bridge

For a vertiport operator at Gimpo or on a Gangnam rooftop planning a 2027 commercial launch, the Incheon Technopark validation run is the closest available Korean site analogue. The environmental load is comparable, the permit envelope is analogous, and the audit-chain methodology is directly portable. The pipeline does not need to be re-engineered for each new site — it needs to be re-calibrated for local species pressure and deterrent-response thresholds, both of which the AVIX-AI BirdThreat configuration layer supports without hardware change.

---

2. Problem Definition — The Audit-Chain Gap in K-UAM Wildlife Management

Korean aviation wildlife management requirements derive from two converging sources. ICAO Doc 9332 establishes the international doctrine for wildlife hazard management, requiring continuous monitoring, documented deterrent activity, and strike-rate trend analysis. Domestically, KAC operational standards and the Korean Aviation Safety (KAS) regulatory framework translate those obligations into vertiport-specific permit conditions. The MOLIT K-UAM Roadmap 2030 projects 200+ vertiports across the Korean peninsula, a significant proportion of which fall within 15 km of EAAF flyway pinch points — tidal flats, estuaries, and reservoir systems that concentrate migrating birds at precisely the altitudes where eVTOL approach and departure paths operate: 0–300 m AGL.

The structural problem is that existing wildlife management programmes at Korean airports rely on human patrol logs, periodic bird-activity surveys, and reactive deterrent deployment. These methods produce documentation that is neither continuous nor machine-verifiable. A patrol log entry timestamped at 07:30 does not document what occurred at 07:45. For a legacy airport with a 3,000 m runway and a 90-second response window, that gap is manageable. For a vertiport with a 30-second pad occupancy cycle and rotor systems that cannot tolerate even sub-100g debris ingestion, it is a certification liability.

The gap is quantifiable: ICAO Doc 9332 recommends a minimum of one wildlife hazard assessment per year at aerodromes below Category I — a standard that produces a single annual document, not a continuous event log. KAC wildlife management reports for FY2023 document patrol frequency but do not publish machine-readable event continuity metrics. No current Korean vertiport operator has a published, auditable, continuous wildlife-detection record. That is the gap AVIX-AI BirdThreat is designed to fill before the certification pressure of 2027 arrives.

---

3. UAM KoreaTech Solution — The 4-Stage Pipeline Decomposed

AVIX-AI BirdThreat (Pillar E) executes four discrete, auditable stages against every entity event detected within a vertiport's ground habitat zone.

Stage 1 — Identify: A Jetson Orin Nano edge node processes sensor input in real time, isolating movement signatures from background clutter. The Orin Nano's 40 TOPS inference throughput enables sub-200 ms entity detection latency without cloud round-trip dependency, which is critical for maintaining pipeline continuity during connectivity degradation — a realistic failure mode at coastal and rooftop Korean sites.

Stage 2 — Classify: A trained convolutional inference model assigns each detected entity a species or species-group label, a size class (small / medium / large), and a flock-density estimate. Classification confidence scores are logged with each event record. This stage is what separates a deterrent pipeline from a motion-triggered alarm: deterrent proportionality requires knowing what you are responding to.

Stage 3 — Respond: The classified threat level triggers a proportional deterrent output — acoustic, visual, or combined — calibrated to avoid habituation. Habituation is the primary failure mode of legacy bird-scaring systems: a species exposed to a uniform deterrent stimulus for more than 72 hours shows measurable approach-distance reduction. AVIX-AI BirdThreat's response layer rotates stimulus parameters within a deterrent envelope, maintaining efficacy across a 30-day operational window without manual reprogramming.

Stage 4 — Monitor: Every event — detection, classification, deterrent trigger, entity exit — is logged with a timestamped HTTP 200 confirmation, producing a machine-readable, gap-free audit record. At Incheon Technopark, all 19 sequential test cycles returned HTTP 200, meaning no event produced a silent failure, a dropped log entry, or an unconfirmed deterrent trigger. Detected animal-class entities are published natively into Anduril Lattice, enabling multi-site vertiport operators to maintain a Common Operating Picture (COP) across their network without bespoke integration work.

The pipeline's provenance discipline — every stage producing a verifiable, chained record — is what makes it a certification asset, not merely an operational tool.

---

4. Strategic Context — Why Korea, Why 2027

The K-UAM Roadmap 2030 (MOLIT, 2023) commits the Korean government to a phased commercial UAM launch, with the first commercial corridor operations targeted for 2027. The roadmap identifies 200+ vertiport sites, a majority of which sit within the EAAF flyway corridor running from the Yellow Sea coast through the Han River basin to the eastern metropolitan fringe. This is not a coincidence — Korean urban density means vertiports will inevitably be sited near the wetland and estuarine habitats that concentrate migratory birds at low altitude.

MOLIT working-group deliberations through 2025–2026 have progressively tightened the environmental compliance annex of draft vertiport certification standards. The emerging consensus is that operators will need to demonstrate continuous, documented wildlife hazard management as a pre-condition for pad activation, not merely a post-incident obligation. That regulatory direction aligns exactly with what the AVIX-AI BirdThreat pipeline produces.

The Anduril Lattice integration layer also positions the pipeline within the dual-use airspace management architecture that Korean defence and civil aviation planners are co-developing for LAANC-equivalent low-altitude airspace. Entities published into Lattice carry structured provenance — sensor origin, classification confidence, deterrent response timestamp — enabling cross-domain handoff between civil vertiport operators and military low-altitude airspace coordinators. For dual-use VCs scoping the 2027 window, that architecture depth is the differentiating factor between a point solution and a platform investment.

KAS Part 25 compatibility of the companion Acoustic Vibration Mat (Pillar E) further reinforces the site certification stack: operators who install the Mat for rooftop acoustic damping at 8–40 Hz and deploy AVIX-AI BirdThreat for wildlife management can present a unified Pillar E compliance package to certification authorities, reducing the per-site documentation burden substantially.

---

5. Forward Outlook

The 19/19 HTTP 200 validation at Incheon Technopark marks the close of the pre-commercial proof phase. The 12-month operational roadmap through mid-2027 has three primary milestones. First, multi-site deployment across the Incheon–Gimpo–Yeouido corridor, building the first Korean vertiport-network wildlife-event dataset at sufficient density to support MOLIT audit review. Second, species-model refinement using Korean-coastal bird populations — specifically the EAAF-listed shorebird assemblages that represent the highest density threat during March–May and August–October migration windows. Third, Lattice COP dashboard integration for network-level operators managing five or more pads simultaneously, enabling automated deterrent-response coordination across adjacent sites to prevent displacement-driven hazard migration.

The 2027 commercial window will reward operators who arrive at certification review with a continuous, machine-auditable wildlife management record. The operators who begin deploying now accumulate that record organically. Those who wait until 2026 to begin procurement face a compressed timeline to demonstrate the operational history that certification authorities will expect.

---

Conclusion

The AVIX-AI BirdThreat 4-stage pipeline — Identify · Classify · Respond · Monitor — is not a wildlife-deterrent product with a data-logging feature. It is a machine-auditable airspace response system that happens to address the wildlife-hazard compliance gap that every Korean vertiport operator will face at the 2027 certification gate. The 19/19 HTTP 200 result at Incheon Technopark is the evidentiary baseline; the Anduril Lattice publish layer is the network architecture that scales it. Operators who deploy the pipeline before the SERP rush becomes a certification rush will be the ones who open their pads on schedule in 2027 — not the ones explaining a documentation gap to a MOLIT inspector.