Evaluating Lost Person Behavior Models

In short: When someone is lost in the wilderness, where should searchers look first? Testing three ways of drawing search maps against hundreds of real cases in the United States, researchers found that simple distance rings around the last known point already cut out over three quarters of the search area, and adding ridges and watersheds did slightly better. Most lost persons did not travel far.

Summary by Kersey Lawrence, Original Wisdom · Updated 9 November 2026

Citation and access

Authors: Elena Sava, Charles Twardy, Robert Koester and Mukul Sonwalkar
Year published: 2016
Journal or source: Transactions in GIS
Full citation: Sava, E., Twardy, C., Koester, R. and Sonwalkar, M. 2016. Evaluating lost person behavior models. Transactions in GIS 20(1): 38-53. (First published online 24 June 2015.)
DOI: 10.1111/tgis.12143 (opens in a new tab)
Access status: Publisher link only

Where to find it (opens the publisher's page in a new tab)

Detailed keywords

Species: Studied: humans (lost persons in wilderness search and rescue incidents, mostly hikers). No animals studied.
Location and habitat: United States. Historical search incidents from New York State (farms, forests, rivers, rolling mountains and lakes; Northeastern Highlands, Erie Drift Plain, Eastern Great Lakes Lowlands and Atlantic Coastal Pine Barrens ecoregions), Arizona (transitional between plains and mountains) and Yosemite National Park, California (a rugged valley between granite peaks in the Southern Sierra Nevada ecoregion). Cost figures are for the US National Park Service.
Topics: wilderness search and rescue (WiSAR); lost person behavior; probability maps; where a lost person is likely to be found; distance from the initial planning point (IPP); distance rings (25%, 50%, 75% and 95%); watersheds and ridges as barriers to travel; combining models; MapScore, a public website for scoring probability maps; search theory; probability density; search costs and time; chance of a successful rescue over time; motion models.
Statistics or models: Euclidean (straight-line) distance ring model from Koester (2008) quartile tables; watershed model counting watershed crossings (US Geological Survey sub-watershed, 12-digit hydrologic unit layer); combined distance-watershed model built on the joint distribution of both; Rossmo’s rank-order scoring metric (R, from -1 to +1, random maps score 0); means with 95% confidence intervals; two-tailed paired t-test; one-way analysis of variance (ANOVA) with F-statistics by region. Maps were built with ArcGIS tools (IGT4SAR, MapSAR, Multiple Ring Buffer, Union).
Data type: Existing data: historical cases from the International Search and Rescue Incident Database (ISRID) and Yosemite National Park data, with reliable start and find coordinates; no collection dates given. Modelling: three probability-map models scored against the actual find locations. Inconsistencies in the paper: (1) the text says 376 cases had reliable coordinates and that “a third” came from each region, but gives 89 New York, 88 Arizona and 199 Yosemite cases; the watershed statistics (Table 2) use 398 cases (102 New York, 122 Arizona, 174 Yosemite) and the combined statistics (Table 3) use 355; (2) the text says the watershed statistics are “shown in Table 1”, but they are in Table 2; (3) the text says about 1 in 17 cases (about 6%) were found three or more watersheds away, but Table 2 gives 19 of 398 (about 5%); (4) the text lists F-values of 11.4 and 8.6 for the Distance and Combined models in that order, but Table 4 gives 8.588 for Distance and 11.366 for Combined; (5) the conclusions say all models did worse in Arizona, while the results report that the regional difference was not statistically significant for the watershed model.

Synopsis

1. What they did

When someone goes missing in the wilderness, searchers start with a probability map: a map of where the lost person is most likely to be found. Most searches are solved within a few hours, but the rest need careful planning, and the chance of a successful rescue drops the longer a search goes on. Many ways of drawing these maps exist, but no one had tested how accurate they are. The authors set out to do that.

A scoring website. They built MapScore, a public website where anyone can upload a probability map for a real historical case and have it scored against where the person was actually found. The score (R) runs from -1 to +1: a blank or random map scores 0 and a perfect map scores 1. It shows how much better a map does than searching at random.
The cases. They used 376 lost person cases with reliable start and find locations, from the International Search and Rescue Incident Database (ISRID) and Yosemite National Park, covering New York, Arizona and Yosemite.
Three models.
Distance rings: the most widely used method. Circles are drawn around the initial planning point (IPP, usually where the person was last seen) holding 25%, 50%, 75% and 95% of past finds, based on Robert Koester’s tables for each type of lost person (for example hikers) and terrain.
Watersheds: this model counts how many ridges (watershed boundaries) lie between the starting point and each area, since ridges are real barriers to travel. It ignores the type of person and the terrain.
Combined: the rings and watersheds laid over each other, with probabilities taken from how often past lost persons were found in each combination.

2. What they found

Distance rings scored 0.78 on average (95% confidence interval 0.74 to 0.82). On average this cut out over 75% of the search area.
Watersheds alone scored clearly lower, at 0.61 (0.57 to 0.65), cutting out about 60% of the search area.
Combined scored best, at 0.81 (0.77 to 0.84), cutting out over 80% of the search area. The gain over distance rings was small but statistically significant.
Where people were found: about half of the lost persons (215 of 398) were found in the same watershed as their starting point, and another 144 one watershed away. Most did not travel far.
By region: all models did best in New York and worst in Arizona, where scores also varied most. The combined model did best of all in New York (0.887). The authors suggest that development and vegetation limit travel in New York, that the New York watersheds were larger, and that the Arizona and Yosemite terrain is more rugged.
Ridges matter: when the starting point lay on a ridge, it was essentially random which side counted as the “same” watershed, which weakened the watershed model.
Cost and time: in 2012 the US National Park Service ran 4,080 search and rescue operations at a total cost of $5.3 million, and costs appear to be rising faster than inflation. The chance of a successful rescue falls over time for both hikers and children aged 4 to 6.

3. What they concluded

The simple distance ring model sets a high bar. Adding even basic terrain information (watersheds) improved it slightly.
New and more detailed models may seem more appealing, but they need to be tested against real cases. MapScore offers the cases, a scoring method and baseline scores to beat. The authors hoped to see models scoring above 0.9 within a year.
Next steps: smoother distance models without sudden jumps at the ring edges; finer watershed maps, since the current ones are too large for search purposes; better handling of starting points on ridges; adding other barriers such as streams and slopes, or travel-cost maps; and, finally, testing models of how lost persons move over time.

6. Other technologies, skills or innovations

MapScore website (mapscore.sarbayes.org) for scoring probability maps against real historical cases, with a public leaderboard and batch uploads.
Geographic Information System (GIS) tools in ArcGIS: IGT4SAR (Integrated Geospatial Tools for Search and Rescue) and MapSAR toolboxes.
US Geological Survey watershed boundaries (12-digit hydrologic units) as a simple stand-in for terrain.

8. Implications for tracking (Original Wisdom's observations)

These are Original Wisdom's observations, not the authors'.

The models in this paper predict where a lost person is likely to be, not the route they took. Tracks and signs found on the ground show what a person actually did, which is the information the authors say is missing when they try to build models of how lost persons move.
Most lost persons in these cases were found close to their starting point, often in the same watershed. For trackers working a search, this suggests where the trail is most likely to lead, and that ridges may act as barriers to travel.

Questions and answers

What did Sava and Twardy et al. (2016) study, and how?

When someone is lost in the wilderness, where should searchers look first? Testing three ways of drawing search maps against hundreds of real cases in the United States, researchers found that simple distance rings around the last known point already cut out over three quarters of the search area, and adding ridges and watersheds did slightly better. Most lost persons did not travel far. Read the full section

What did Sava and Twardy et al. (2016) find?

Distance rings scored 0.78 on average (95% confidence interval 0.74 to 0.82). On average this cut out over 75% of the search area. Watersheds alone scored clearly lower, at 0.61 (0.57 to 0.65), cutting out about 60% of the search area. Read the full section

What did Sava and Twardy et al. (2016) conclude?

The simple distance ring model sets a high bar. Adding even basic terrain information (watersheds) improved it slightly. New and more detailed models may seem more appealing, but they need to be tested against real cases. Read the full section

Cite this summary

Lawrence, K. 2026. Summary of: Evaluating Lost Person Behavior Models (Sava and Twardy et al., 2016). Original Wisdom, Tracking Literature. https://www.originalwisdom.com/tracking-literature/2016-sava-twardy-evaluating-lost-person-behavior/

For the paper itself, cite the full citation above.

← Back to Tracking Literature