AI · Model card

Model Card: SKYSWEEP FOD Detector v2.1

About this sample

Sources checkedTotals verified

Audience

Airfield operations leads deciding whether to rely on SKYSWEEP, and the review board that approves its use.

Document type

Model card, mapped to the five DoD AI Ethical Principles (2020) and the four functions of the NIST AI RMF 1.0 (NIST AI 100-1).

How this sample was made and checked

Approach

Performance is reported per operating condition rather than as one average, so the fog limitation leads straight to its mitigation, and each row of the alignment tables points to the section that supports it.

How it was checked

The principle and function names match the DoD AI Ethical Principles and NIST AI 100-1, and a test checks the names and re-sums the frame counts on every change.

Sample content

SKYSWEEP is a fictional system. All metrics and datasets are illustrative.

Field	Value
Model name	SKYSWEEP FOD Detector
Version	2.1.0
Model type	Single-stage object detector (convolutional backbone)
Input	1920 × 1080 RGB frames from fixed runway-edge cameras
Output	Bounding boxes, object class, and confidence score (0.0–1.0)
Owner	Airfield Systems AI Team (fictional)
Last reviewed	2026-09-30

1. Intended Use

SKYSWEEP flags possible foreign object debris (FOD) on runways and taxiways so that airfield personnel can inspect and clear it.

Intended users: Airfield management and FOD-walk crews.

Decision authority: SKYSWEEP is a decision-support tool. A trained human confirms every detection before any runway closure or clearance action.

Deactivation: Operators can disable SKYSWEEP at any time; scheduled FOD walks continue as normal.

Out-of-scope uses

Do not use SKYSWEEP to:

• Certify a runway as clear without a human inspection
• Detect wildlife, personnel, or vehicles (the model was not trained for these classes)
• Process imagery from moving or airborne cameras

2. Performance

Evaluated on a held-out test set of 12,400 frames from three airfields not seen in training.

Condition	Recall	Precision	Frames
Daylight, dry	0.96	0.91	6,800
Daylight, wet	0.92	0.86	2,100
Night (IR-assisted)	0.88	0.83	2,700
Fog / visibility < 800 m	0.71	0.78	800

Known limitation

Recall drops to 0.71 in fog. In low-visibility conditions, operators must perform a manual FOD walk regardless of model output.

Detection threshold: 0.45 confidence. Lowering the threshold increases recall but roughly doubles the false-alarm rate.

3. Training Data

Attribute	Description
Source	310,000 labeled frames from 9 airfields (fictional)
Classes	Metal hardware, tire fragments, plastic, stones > 2 cm, tools
Labeling	Two independent annotators per frame; disagreements resolved by a third reviewer
Known gaps	Snow-covered surfaces are underrepresented (< 1% of frames)

4. Limitations and Failure Modes

Failure mode	Effect	Mitigation
Debris smaller than 2 cm	Missed detection	Scheduled manual FOD walks remain required
Painted runway markings in low sun	False alarms	Operator review; per-camera masking
Camera lens contamination	Degraded recall	Automated image-quality check halts inference and alerts the operator
Dataset shift (new pavement type)	Unknown	Re-validate before deployment at any new airfield

5. Responsible AI Alignment

This section maps the model to the DoD AI Ethical Principles and the NIST Artificial Intelligence Risk Management Framework (AI RMF 1.0, January 2023).

DoD principle	How it is addressed
Responsible	Humans confirm every detection; the model cannot close or clear a runway
Equitable	Test set spans three unseen airfields and four lighting/weather conditions, with results reported per condition (Section 2) to expose uneven performance
Traceable	Training data lineage, labeling procedures, and model versions are recorded
Reliable	Performance reported per operating condition, not as a single average
Governable	Operators can disable the model (Section 1); the image-quality check halts inference and alerts the operator (Section 4)

NIST function	Artifact
Govern	Model owner and last-review date listed in this card; re-validation required before deployment at any new airfield (Section 4)
Map	Intended and out-of-scope uses (Section 1)
Measure	Condition-level metrics (Section 2)
Manage	Failure modes and mitigations (Section 4)

6. Revision History

Version	Date	Change
2.1.0	2026-09-30	Added night (IR-assisted) evaluation; documented fog limitation
2.0.0	2026-05-12	Retrained on 3 additional airfields
1.0.0	2025-11-03	Initial release