The Science Behind SAROS
Search decisions carry real consequences, so a planning tool must be able to answer two questions plainly: where do its numbers come from, and would it give the same answer twice? This page sets out the models SAROS implements, the published sources they derive from, how the computation is kept reproducible and auditable, and how the results have been tested against real incidents.
Every behavioural distribution, leeway coefficient and detection model traces to a named published source — Koester, the Allen leeway field studies, UK Mountain Rescue research, Grampian Police case data, the IAMSAR Manual. The citations live in the source code beside the numbers they justify.
Monte Carlo does not mean arbitrary. Drift and Aero ensembles run from a recorded random seed: the same inputs and seed reproduce the same particles, the same contours, the same plan — reviewable months later in a debrief or inquiry.
Live drift trials with a police marine unit, 21 real incidents from the published Koester casebook re-run retrospectively, IAMSAR sweep-width tables verified cell-by-cell, and ~3,800 automated tests executed before every release.
SAROS is decision support for qualified operators, not an oracle. Every output declares its assumptions, every simplification is stated, and residual probability outside the searched area is always shown. Absence of probability is never treated as evidence of absence.
The International Aeronautical and Maritime SAR Manual is the operational standard. SAROS implements it directly rather than approximating it.
The IAMSAR Vol III uncorrected sweep width tables for merchant vessels, helicopters and fixed-wing aircraft (Tables 3-19 to 3-21) are carried as data with their source recorded — and an automated test validates every cell against the published values, including all weather-correction combinations. Wc = Wu × fw (IAMSAR corrected sweep width)
Probability of detection uses the IAMSAR exponential (random-search) detection model from a single shared implementation across all three domains — the same POD mathematics governs a vessel sweep, a hillside segment and an air search area. POD = 1 − e−C, C = Wc / S (IAMSAR coverage)
Parallel track, expanding square and sector search generated per IAMSAR Vol III Chapter 4 geometry — sized to the particle spread at commence-search time, with waypoints, bearings, leg times and estimated duration on an exportable search card. Sector legs R, R, 2R, R, 2R, R, R (9R total) · Expanding square S, S, 2S, 2S…
The IAMSAR sweep width tables were measured for human lookouts. A modern EO/IR camera — with or without an AI detection system — has no published entry in those tables, yet sweep width is the single number all search planning runs on. SAROS closes that gap by extending classical search theory rather than replacing it.
Instead of asking "did the system see the target?", the framework asks "was the system given a fair opportunity to see the target — and did it succeed?" Detection performance is counted over glimpse opportunities: windows in which the target was genuinely visible for long enough to be detectable. The result is a detection probability that is testable, repeatable and comparable across sensors — properties a human lookout's performance never had. SAROS also verifies the aircraft spends enough frames on target at the planned speed for the detector to work, and flags the plan when it doesn't.
Noticing something is not the same as knowing it is a person and not a lobster pot. The framework keeps the two apart, producing a pair of sweep widths: a wider one for detection alone and a narrower one for detection plus recognition (POD+R). The coordinator chooses explicitly between covering more water and holding higher confidence in what the sensor reports — a planning lever that simply does not exist with traditional table lookups.
Sweep width has meant the same thing since Koopman's wartime work: the area under the curve of detection probability against distance off track. SAROS derives that curve from the sensor's own physics — resolution, optics, geometry, visibility — and integrates it exactly as classical theory prescribes, so the result plugs directly into the IAMSAR coverage and POD machinery above with no change to doctrine or workflow. W = 2 ∫ POD(r) dr (Koopman, Search and Screening, 1946–56)
A camera can only search the strip of water its field of view actually covers. SAROS computes the usable corridor from the sensor's true geometry — and caps the sweep width accordingly, rather than quoting the theoretical detection range of a perfectly aimed pixel. Where viewing geometry degrades (grazing angles at long range), the result is flagged rather than silently trusted. An honest, smaller number that a search plan can rely on beats an impressive one that it cannot.
The evaluation framework behind this capability — standardised detection appraisal and the POD+R measure — is published research from IASARC — the International Association of Search and Rescue Coordinators — developed so that any sensor, human or machine, can be compared within the same probabilistic framework. Sweep widths computed in SAROS are physics-derived; per-sensor detection probabilities are coordinator-editable and should come from standardised appraisal of the specific sensor where available.
How a drifting object moves is one of the most-studied problems in SAR science. SAROS implements the leeway approach established by the US Coast Guard field-study programme and codified in the IAMSAR Manual.
A drifting object moves with the water it sits in, plus a wind-driven component — leeway — that depends on how much of the object faces the wind above and below the waterline. Decades of field experiments — principally Allen & Plourde (1999) and Allen (2005), refined by Breivik et al. (2011) — measured this relationship for real SAR objects by instrumenting them and recording drift against measured wind. SAROS decomposes leeway into downwind and crosswind components relative to the wind, on top of tidal and ocean-current advection.
SAROS carries a versioned object-class library — 15 classes in four groups: persons in water in different postures, ballasted and unballasted liferafts, small craft and kayaks, surfboards and debris — each with its empirically derived downwind slope and offset, leeway uncertainty, crosswind divergence angle δ and jibing rate. One class, the sit-on-top kayak, is additionally calibrated directly against a physical drift trial SAROS ran with NZ Coastguard in the Hauraki Gulf (15.1 km drift over 5.8 hours).
Particles deployed across the Last Known Position uncertainty, from a recorded random seed
Each particle advected by measured wind, tide and current, with field-study leeway coefficients for the object class
Probability contours computed from where the ensemble actually went — priority regions, not a single guess
Two live targets — a kayak and a person in water — drifted 15 km and 7.5 km over 8 hours of Auckland tide and offshore wind. Both were recovered inside the SAROS-planned search area, 0.7 and 0.27 miles from the predicted positions.
Where a lost person goes is not random. It follows patterns measured across thousands of resolved incidents — and that measured behaviour, not intuition, is what SAROS simulates.
The published percentile tables define, for each subject category, how far subjects are found from the point last seen. SAROS fits a log-normal distribution to those percentiles and runs thousands of simulated journeys against the real terrain.
Twenty-one real search incidents documented in Koester's published casebook — dementia, autism, children, hikers, hunters, despondent subjects — were re-run through SAROS from the original initial planning point. Every case is scored, including the five subjects found at the planning point itself; the three finds beyond P95 are published by name rather than removed.
A hiker in the Cairngorms does not behave like a hiker in the Rockies. SAROS is one of the few tools that carries UK-calibrated tables — Perkins & Roberts and the Grampian dataset — alongside the US ISRID-derived values, with a documented fallback chain when a category lacks regional data.
For an overdue aircraft the physics is unforgiving and therefore tractable: a gliding aircraft trades height for distance at a known ratio, drifting with the wind as it descends.
The Monte Carlo layer then perturbs what is genuinely uncertain — position, altitude, heading, glide performance, wind — across ensembles of up to 20,000 particles, producing probability heatmaps and nested containment contours. The probability accounting is exact by construction: cell masses plus the residual outside the grid sum to one; nothing is renormalised away.
The glide engine is verified against a benchmark set of reference cases — including the specification's worked examples — and must reproduce them within 1% on every test run. The Monte Carlo layer is seeded: if no seed is supplied, one is generated and recorded with the results, so any ensemble can be re-run identically.
"Would it give the same answer twice?" — the property that turns a model into defensible decision support.
Drift and Aero Monte Carlo ensembles run from an explicit random seed carried with the results. Identical inputs and seed produce identical particles, contours and search areas — asserted by automated tests on every release. Land search validation is statistical, measured by percentile containment across the published case benchmark.
Every run is stamped with the model version and the object-class library version that produced it. Case assumptions are versioned append-only — author, time, reason and a full snapshot — so a plan can be reconstructed exactly as it stood at any decision point.
Evidence records (radar, ADS-B, witness, debris…) carry position, time, uncertainty and confidence. Excluding evidence requires a recorded reason — it is flagged out, never deleted. Environmental data provenance records which provider supplied each forcing field and any fallback taken.
SAROS is decision support and is engineered to stay that way: acknowledgement gates on ground-frame patterns in set, nearshore model limits and residual risk; assumptions printed on every output; a full audit trail on all operations; and ~3,800 automated tests run before every release.
Four layers, from physics to the real world.
Physics engines reproduce published worked examples and golden benchmark cases within stated tolerance (glide: 1%) on every automated test run.
IAMSAR sweep-width and weather-correction tables validated cell-by-cell against the manual. Behavioural tables carry their bibliographic source in code.
21 published Koester casebook incidents re-run from the original planning point: 18 of 21 finds (86%) inside the P95 search area, against a pass target of 80% declared before the run.
Instrumented drift targets with the NZ Police Marine Unit: both recovered inside the SAROS search area. Object-class coefficients calibrated against the measured tracks.
A pre-registered population of 100 real ocean drift tracks — selection rules and acceptance thresholds committed in writing before any data was drawn, spanning both hemispheres — plus the fully documented Ouzo person-in-water case (MAIB 7/2007). Ninety-nine of one hundred inside threshold; median error 0.18 NM over 48-hour drifts (0.48 NM across all durations to five days), nine in ten within 1.74 NM. Rejected and excluded tracks are published with the same prominence as the passes. Runs on every release. Full results →
No international SAR authority has ever published a pass mark for drift-prediction accuracy — not the IMO, not the IAMSAR Manual, not any national agency. The question has a governing precedent instead. CASP, the US Coast Guard's computerised search planning system, served for over three decades without such a threshold ever existing. When SAROPS replaced it in 2007, the case for fielding rested on three demonstrations, not a number: the components were validated against field science; the whole performed at least as well as the incumbent; and the uncertainty was characterised, so planners knew how far to trust it.
SAROS answers the same three tests, with more of the evidence published. Its leeway physics closes exactly against the US Coast Guard field-trial literature; it has been benchmarked head-to-head against the leading open implementation of the same science; and a pre-registered population of one hundred real drift tracks characterises both its accuracy (median error 0.18 NM over 48-hour drifts) and the honesty of its stated uncertainty (spread-to-error ratio 0.97 against an ideal of 1.0). How the uncertainty is measured →
SAROS's own validation reports — the Validation Test Report and accuracy datasheet for the real-outcome golden set, the NZ Police live trial results, the 21-case Koester retrospective benchmark, the sweep-width verification suite, and a head-to-head benchmarking study against the leading open implementation of the same leeway science — are available to prospective customers and professional reviewers on request.
A model you can trust is one that tells you where it stops. These limitations are documented in the SAROS source code and on the relevant outputs; the roadmap items among them are tracked openly.
| Decision support | SAROS assists qualified human operators. It does not replace SAR doctrine, training or command judgement, and its terms of use make acknowledging this a condition of access. |
| Conditional probability | All displayed probability is conditional on the modelled scenario set. The subject may lie outside every modelled scenario — residual probability is always displayed, never hidden. |
| Aero performance data | Aircraft profiles are generic engineering values, not type-certified data — stated on every output. Retrospective validation against published accident investigations is a declared roadmap requirement. |
| Behavioural coverage | Distance tables for all categories are published-source; behaviour-state weights for some categories still use general Koester heuristics pending full per-category literature review — tracked as an open backlog item. |
| Weather uncertainty | Forecast wind and current are applied as provided; forecast-error propagation through the ensemble is a roadmap item. Provider and fallback provenance are recorded on every run. |
| AI-sensor sweep widths | Sensor sweep widths are derived from sensor physics and stated detection assumptions. Default detection probabilities are representative values; each sensor's true figures should come from standardised appraisal, and every assumption is printed on the Coordinator Briefing so the basis of the plan is never hidden. |
We welcome scrutiny from SAR professionals, scientific advisors and procurement teams — including live walkthroughs of the models and validation evidence.