🔔 Subscribe to ShipUniverse Weekly →
The Coast Guard Wants AI Searching for People Lost at Sea: Can a Machine See Through Fog Better Than a Human?
The Coast Guard is preparing to measure automated detection against the same problem that defines every search: how likely is the searcher to see a small target before passing it? Fog, darkness, rain and rough seas make the answer much harder than an AI accuracy score.
The algorithm searches.
A person floating in the ocean can present little more than a head, shoulders and flotation device above the surface. Add whitecaps, spray, glare, darkness or fog and the search target can disappear into thousands of square miles of moving visual clutter.
The Coast Guard's next experiment is not simply asking whether artificial intelligence can recognize that person in a camera image. It wants to know whether an automated detection system can search continuously enough, far enough and reliably enough that its performance can be inserted into the mathematics used to plan a real rescue.
That distinction matters. An AI detector can score 98% on a curated image set and still be useless if its camera sees only a narrow strip of water, if dense fog destroys target contrast or if whitecaps trigger so many alarms that the aircrew spends the mission investigating false targets.
The Coast Guard is not buying an autonomous rescue officer
No procurement contract has been awarded
A Cooperative Research and Development Agreement lets government and a private participant share personnel, equipment and research resources. The Federal Register explicitly states that this type of agreement is not a procurement contract.
The Coast Guard wants mature, commercially available technology rather than an early laboratory prototype.
The system needs to move between operational test environments instead of requiring a custom laboratory installation.
Automated detection must work locally, an important requirement for aircraft, cutters and remote maritime operations.
The objective is to analyze Coast Guard sensor feeds rather than lock SAR performance to one proprietary camera.
The real test is not accuracy. It is lateral range.
Detection probability changes with distance
A searcher is highly likely to detect a target close to the aircraft or vessel track and progressively less likely as lateral distance grows. That relationship is what the Coast Guard wants measured for automated detection.
The target occupies more pixels, presents more sensor signal and should have the highest detection probability.
A person becomes a smaller target while atmospheric attenuation and sea clutter consume more of the signal.
Fog, rain, darkness and sea state alter the effective curve differently for visible, infrared and other sensors.
Once the curve is measured, search planners can determine how tightly tracks need to be spaced to achieve useful coverage.
Can AI see through fog better than a human?
Sometimes the system can. The reason is the sensor, not artificial intelligence itself.
Performance falls sharply without illumination.
Water droplets scatter visible light and can destroy useful contrast.
Can maintain continuous attention and enhance weak patterns, but cannot restore information absent from the image.
Major advantage because thermal cameras do not depend on visible illumination.
Certain infrared bands can outperform visible light, especially in lighter fog, but dense fog and rain still reduce range substantially.
Continuously searches thermal contrast for human-sized signatures that an operator could overlook.
Essentially irrelevant to radar performance.
Radar wavelengths can retain significant all-weather utility, although a human body barely above the water is an extremely difficult target amid sea clutter.
Can correlate weak or intermittent tracks with other sensor detections rather than treating radar as a standalone answer.
Measures many spectral bands instead of ordinary red-green-blue imagery.
Wide Area Maritime Object Recognition is evaluating commercial hyperspectral sensing for detection and classification in poor visibility.
Classifies spectral signatures across data that would be impossible for a lookout to inspect manually in real time.
The advantage changes with the weather
This is the latest step in a longer Coast Guard experiment
Make people easier to see
The Enhanced Person in the Water Detection program evaluated ways to increase conspicuity and reduce search burden.
Test AI on optical feeds
A Coast Guard CRADA sought methods to compare autonomous casualty detection against accepted SAR standards.
Move AI to operational sensors
Computer vision flew from National Security Cutters while unmanned SAR demonstrations tested person-in-the-water recovery.
Quantify machine search
The new effort aims to generate a coverage factor and data-driven Lateral Range Curve suitable for SAR planning.
A small improvement in detection compounds across enormous search areas
Multiple Coast Guard air and surface assets searched this area over roughly 24 hours after the fishing vessel casualty off Massachusetts.
Coast Guard and local responders accumulated more than 80 combined search hours for one missing person.
Search efforts covered more than 1,090 nautical miles with aircraft and ships despite favorable visual conditions.
The likely future is machine search, human rescue judgment
Multiple feeds can be monitored simultaneously without asking one aircrew member to divide attention between them.
A person repeatedly hidden by wave crests can potentially be associated across multiple frames instead of rediscovered each time.
Thermal, visible and other detections can be correlated into one candidate track.
Machine performance can be recorded and later used to build empirical detection curves.
A partially submerged cooler, bird, whitecap, debris field and human body can produce similar transient signatures.
Search coordinators consider drift, survival time, new evidence, asset endurance and multiple competing scenarios.
Finding a candidate is only the beginning. Aircrews and boat crews still have to verify, approach and recover the person safely.
Model confidence cannot replace operational responsibility for continuing, redirecting or ultimately suspending a search.
Six ways the AI searcher can fail
Foam, spray, floating debris and reflections create transient shapes and temperature patterns.
Rough water can hide nearly the entire body between waves, leaving only brief detection opportunities.
Infrared performance depends on target-to-background temperature difference as well as atmospheric transmission.
Dense water droplets attenuate both visible and infrared energy. Better classification cannot overcome zero signal.
A system that detects everything but generates hundreds of irrelevant alerts can reduce rather than increase useful search coverage.
Different cameras, altitudes, lenses, sea states, climates and survival equipment can move performance away from laboratory validation.
AI Search Coverage Stress Test
Model how an automated sensor could change search coverage. The simulator uses a simplified coverage relationship to compare a conventional visual search with an AI-assisted sensor. It deliberately penalizes false alerts because every unnecessary investigation consumes real search time.
Does AI actually widen the search?
Enter search area, speed, time, effective sweep width and environmental degradation. The model estimates coverage and probability of detection for screening purposes.
Under these assumptions the automated sensor meaningfully increases effective search coverage without overwhelming the operator with false alerts.
The system becomes operationally important only when its measured range and persistence expand effective coverage after environmental degradation and false-alert workload are included.
The decisive experiment is not whether AI finds the mannequin
Put the target farther from the search track. Change the life jacket color. Add whitecaps. Turn off the sun. Reduce thermal contrast. Add rain. Move the camera to a different aircraft. Change altitude. Partially hide the person behind wave crests. Then measure every detection and every miss.
Do the same thing often enough to produce a curve showing detection probability by range and conditions. Record the false alarms too. If that curve is wider than the human-search curve without creating an unmanageable alert burden, automated detection has done something operationally important.
It has not replaced the rescue crew. It has changed how much ocean that crew can search with confidence.