# Unsupervised recognition of individual tigers and snow leopards from their footprints

> Can a computer tell which tiger left a footprint without being told who walked where? Using hind prints from captive tigers and snow leopards laid in sand, this early study found that a Bayesian program sorted prints into the right animals almost perfectly for tigers and well for snow leopards, while a neural network method failed. The author sets out what still has to be solved before it can be used on wild tracks.

Summary by Kersey Lawrence, Original Wisdom. Updated 9 November 2026. Page: https://www.originalwisdom.com/tracking-literature/1998-riordan-unsupervised-recognition-individual-tigers/

## Citation and access

- **Authors:** P. Riordan
- **Year published:** 1998
- **Journal or source:** Animal Conservation
- **Full citation:** Riordan, P. 1998. Unsupervised recognition of individual tigers and snow leopards from their footprints. Animal Conservation 1(4): 253–262.
- **DOI:** 10.1111/j.1469-1795.1998.tb00036.x
- **Access status:** Publisher link only
- **Where to find it:** https://doi.org/10.1111/j.1469-1795.1998.tb00036.x

## Tags

- **Species:** Mammals > Snow leopard; Mammals > Tiger
- **Tracking:** Individual identification; Track measurements; Track photographs; Tracking; Tracks
- **Methods and technology:** Bayesian statistics; Descriptive statistics; Machine learning; Pearson correlation
- **Data type:** New field data
- **Observer qualification:** Not reported
- **Discipline:** Wildlife tracking
- **Year published:** 1998

## Detailed keywords

- **Species:** Studied: tiger (Panthera tigris), 10 captive animals: 7 Bengal tigers (P. t. tigris) and 3 Siberian tigers (P. t. altaica); and snow leopard (Panthera uncia), 6 captive animals. Mountain lion (Felis concolor) is discussed from earlier footprint studies.
- **Location and habitat:** Port Lympne Zoo (the paper does not give its location; the author was at Manchester Metropolitan University, Manchester, UK). Footprints made in enclosures on patches of fine builder’s sand about 2 cm deep.
- **Topics:** identifying individual animals from their footprints; unsupervised classification (grouping prints with no prior knowledge of which animal made them); hind footprints; pugmark tracing on acetate sheets versus photographs; 27 footprint measurements (18 lengths and 9 angles); effects of substrate, slope, speed and different fieldworkers; sand traps for mark-recapture censuses; tiger census methods.
- **Statistics or models:** Kohonen self-organizing map (SOM), a neural network method, with map sizes from 3×3 to 50×50; AutoClass, a Bayesian classification method developed at NASA; principal components analysis (PCA), 10 components keeping at least 95% of the variation; Pearson correlation between measurements; coefficient of variation; normalized variable influence. Inconsistencies in the paper: (1) Table 1 marks Vosthok, Zeyan and Zeyna as Siberian tigers, yet the Discussion refers to a separate classification of “Bengal tigers” in which Nari formed a small class, and that classification is not shown. (2) Citing Smallwood and Fitzhugh (1993), the Discussion says slope affected main pad width and also had the smallest effect on main pad width. (3) The abstract says larger maps did not substantially improve the SOM, while the Results say larger maps gave more separation (but still no accurate classification).
- **Data type:** New field data from captive animals: at least 10 hind footprints per foot for each animal, by acetate tracing for all animals and by photographs for a subset of tigers. Collection dates are not given (received 2 January 1998).

## Synopsis

### 1. What they did

Researchers studying wild tigers have long tried to sort footprint tracings into individual animals by eye, and this has been criticized as subjective. Earlier work on mountain lions showed that tracks of known individuals could be told apart with statistics, but in the field nobody knows in advance which animal made which tracks. The author therefore tested two “unsupervised” methods, which group footprints into clusters on their own, with no labels for which animal made them.

- The animals. 10 tigers (7 Bengal and 3 Siberian) and 6 snow leopards at Port Lympne Zoo.

- The footprints. Fine sand was spread about 2 cm deep in the enclosures, and the animals walked over it one at a time, so each track set came from a known animal. Only hind footprints made at a normal walk or trot were used, because the hind feet often land on top of the front prints. At least 10 prints per hind foot were collected for each animal, traced onto acetate sheets while kneeling directly over the print. For a subset of tigers the same feet were also photographed from directly above, with a low flash to show depth and a scale ruler.

- The measurements. 27 measurements per print: 18 lengths (to 0.5 mm) and 9 angles (to 0.5 degrees), covering the main pad, the toes and the spaces between them.

- The tests. Left and right hind feet were analysed separately. The prints were fed to a Kohonen self-organizing map (SOM), a neural network method, in several map sizes, and to AutoClass, a Bayesian method developed at NASA. Both were also run on 10 principal components (combined variables) instead of the raw measurements. The author then checked whether the clusters matched the real animals.

### 2. What they found

Self-organizing map

- The SOM did not separate individual tigers or snow leopards well enough to identify them. Prints from the same animal tended to lie close together, but not clearly apart from other animals.

- Bigger maps gave a little more separation; square and rectangular maps did equally well.

- Using principal components slightly improved separation for tigers but added scatter, and did not help for snow leopards.

- Tracings and photographs of the same tiger’s foot did not land close enough together to treat the two methods as equivalent.

AutoClass, tigers

- Left hind feet: 149 of 150 prints were placed correctly; one print of Zeyan was grouped with Zeyna. All prints had class probabilities above 0.95.

- Right hind feet: 2 of 145 prints were misplaced (one of Spotty with Pindi, one of Nari with Chambal), with probabilities above 0.94.

- Prints from tracings and photographs of the same animal formed separate groups.

- Including or leaving out the Siberian tigers made no difference to accuracy.

- No single measurement or set of measurements was consistently most important. Angles mattered most for left feet; the distances from the toes to the main pad mattered most for right feet.

AutoClass, snow leopards

- Left hind feet: 59 of 60 prints were placed correctly; one print of Sitar sat almost evenly between two classes (probabilities 0.57 and 0.42).

- Right hind feet were classified less well: prints of Atheni-Moon and Messalina were mixed together, and some of both were grouped with Atheni. Read without knowing the answer, the result would suggest five animals instead of six.

- Using principal components in AutoClass made classification worse for both species.

Why the methods differed

Footprint measurements were noisy: they were correlated with each other, and their variability differed from animal to animal and was greatest for the small angles. The author suggests AutoClass spread this noise across its models, while the SOM could not cope with it.

### 3. What they concluded

- Unknown individual tigers and snow leopards can be accurately identified from their footprints with unsupervised learning (AutoClass), but not with the self-organizing map.

- These results come from one researcher working under standardized conditions. In the field, substrate, slope, animal speed and differences between fieldworkers could all affect the prints, and all need addressing before field use. Speed can be judged from the spacing of prints in a track set, and sand traps or other artificial surfaces could provide prints for mark-recapture censuses.

- In the field, misclassifications will not be obvious. Collecting as many prints as possible from each track set helps, because all prints in a set come from one animal; stray prints placed outside their set’s main class can be ignored. Small classes need care, since they may be a real extra animal, such as one at the edge of the study area or dispersing through it.

- Accuracy may drop as more animals are included. The method could work for other animals with complex footprints, but may not suit small animals, where measurement error could exceed differences between individuals. Further testing in the field is needed.

### 4. How tracking was or was not used

Footprints were the whole basis of the study. Tracks were laid in sand by known captive animals, and only hind prints made at a walk or trot were taken. The paper assumes that an experienced tracker or field biologist can tell front from hind prints and left from right in the field. No trailing was involved.

### 5. Tracker or observer certification

Not reported. All footprints were collected and measured by the author alone, which the paper notes as a limitation; no tracking qualification is described. The author warns that tracings by different fieldworkers may not match.

### 6. Other technologies, skills or innovations

- Acetate tracing on a picture frame placed over the print without touching it, and 35 mm photographs taken straight down with a low-angle flash and a scale ruler.

- Two freely available unsupervised classification programs: the Kohonen self-organizing map (SOM_PAK, Helsinki University of Technology) and AutoClass (NASA).

- Principal components analysis to remove correlation between measurements.

### 7. Implications for conservation (from the paper)

- A reliable, systematic way of identifying individual tigers from footprints has been called for, and this method, with further research, could provide one. It could be particularly important in tiger conservation.

- Any animal with sufficiently complex footprints could potentially be studied this way, so it could be a significant tool in animal ecology and conservation.

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

_These are Original Wisdom's observations, not the authors'._

- The method depends on tracking skill before any computer is used: choosing hind prints, knowing left from right, and judging gait. In the field, keeping prints grouped by track set is what lets misclassified prints be spotted.

- The study shows how much a footprint can vary within one animal, even on even sand, which is worth bearing in mind when identifying individuals from a single print.

## Cite this summary

Lawrence, K. 2026. Summary of: Unsupervised recognition of individual tigers and snow leopards from their footprints (Riordan, 1998). Original Wisdom, Tracking Literature. https://www.originalwisdom.com/tracking-literature/1998-riordan-unsupervised-recognition-individual-tigers/