How Mathematics Improves The World | Counting Wild Animals Without Seeing Every Animal
Imagine trying to count every leopard in a forest.
You cannot line them up.
You cannot ask them to stay still.
You cannot even assume they want to be seen.
Some animals avoid roads.
Some move at night.
Some live in dense vegetation.
Some are photographed repeatedly while others never appear in a camera trap at all.
If you simply count what you see, you have counted sightings.
You have not necessarily counted animals.
The gap between those two statements is where ecological statistics lives.
Wildlife scientists do not solve the problem by pretending the forest is fully observable.
They model the fact that it is not.
That is one of Mathematics’ most useful habits:
When observation is incomplete, estimate not only the population—but also the probability of observing it.
Quick Read
Wildlife population estimation is difficult because animals are not perfectly detectable. A survey can visit a site containing ten animals and see only three. If detection probability is ignored, abundance can be badly underestimated. Capture–mark–recapture methods address this by repeatedly sampling a population and using the fraction of previously marked animals that reappear to infer how many unobserved individuals are likely to exist.
The simplest two-sample estimator follows an intuitive proportion. Capture and mark n1 animals. Release them. Later capture n2 animals, of which m2 are marked. If marked animals have mixed back into a closed population and every individual is equally catchable, then the marked fraction in the second sample should roughly equal the marked fraction in the full population:
m2/n2 ≈ n1/N
so N ≈ n1n2/m2.
Real ecology is more complicated. Animals differ in catchability. Some enter or leave the study area. Some die. Marks can be lost. Capturing an animal may change its future behaviour. Modern capture-recapture models therefore estimate survival, movement, temporary emigration and individual heterogeneity using maximum likelihood, Bayesian inference and hierarchical models. USGS describes capture-recapture as a major framework for estimating wildlife abundance, survival and state transitions under imperfect detection.
One-sentence answer: Mathematics improves the world by making incomplete wildlife observations scientifically useful, separating the number of animals seen from the number likely present and giving conservation decisions a quantified estimate of uncertainty rather than a confident guess.
A Count Is Not a Population
Walk one kilometre through a forest and see twelve deer.
How many deer live there?
Twelve?
Probably not.
You may have missed animals behind vegetation.
Some may have moved before you arrived.
Some may be more active at another time of day.
Some species freeze when approached.
Others flee.
The raw count equals:
population × availability × detectability
in a simplified conceptual sense.
To recover population, we need information about the missing probabilities.
Detection Probability: The Animal Can Exist and Still Be Absent From the Dataset
Detection probability is the chance an individual is observed given that it is present and available for detection.
This sounds like a technical correction.
It changes the whole science.
Suppose two forests each contain 100 birds.
Forest A is open and detection probability is 0.8.
Expected count: around 80.
Forest B is dense and detection probability is 0.3.
Expected count: around 30.
A naive survey concludes Forest A has almost three times as many birds.
The true populations are equal.
Habitat changed visibility, not abundance.
Without modelling detection, ecological comparison can mistake observability for biology.
Capture–Mark–Recapture: Use Repetition to Estimate the Missing Animals
The basic method is elegant.
- Capture a sample.
- Mark individuals uniquely or as a group.
- Release them.
- Allow them to mix back into the population.
- Capture another sample.
- Count how many are already marked.
If the second sample contains many marked animals, the total population is probably not enormous.
If marked animals are rarely recaptured, the population is probably larger.
Recapture frequency becomes evidence about all the animals we did not see.
The Lincoln–Petersen Estimator
Suppose we capture 40 turtles and mark them.
Later we capture 50 turtles.
Ten are marked.
The marked fraction in the second sample is 10/50 = 0.2.
If the marked animals are representative of the whole population, the original 40 marked turtles should be about 20% of the full population.
So:
N ≈ 40 / 0.2 = 200
Equivalent formula:
N̂ = n1n2/m2
One proportion reveals the unseen majority.
Why the Simple Estimator Can Be Biased
If only one marked animal is recaptured, the denominator is tiny and the population estimate becomes huge.
If zero are recaptured, the basic formula breaks entirely.
Small samples therefore produce strong bias and unstable estimates.
A common Chapman correction modifies the estimator:
N̂ = ((n1 + 1)(n2 + 1)/(m2 + 1)) − 1
This reduces small-sample bias under the model’s assumptions.
The lesson is broader than wildlife.
An intuitive estimator can be approximately right and statistically biased.
Mathematics studies not only the answer, but how the answer behaves across repeated samples.
Assumption 1: The Population Is Closed
The simple two-sample model assumes no births, deaths, immigration or emigration between samples.
That is population closure.
For a short trapping session in an isolated pond, closure may be plausible.
For migratory birds sampled months apart, it is not.
If new unmarked animals enter between samples, the marked fraction falls and the estimator interprets the change as a larger original population.
The model becomes biased because biology violated the assumption.
Assumption 2: Marks Are Not Lost
A marked animal that loses its tag returns as apparently unmarked.
The recapture fraction falls.
Population estimate rises.
Marks must therefore be durable and identifiable.
Modern studies may use:
- metal or plastic tags;
- bands;
- PIT tags;
- radio transmitters;
- natural coat patterns;
- photo-identification;
- DNA genotypes.
USGS notes that capture-recapture has evolved toward non-invasive marking through natural markings and DNA collected from sources such as faeces.
Assumption 3: Marking Does Not Change Catchability
An animal captured once may become trap-shy.
It learned.
Or it may become trap-happy because bait was rewarding.
Then marked animals have different future detection probability from unmarked animals.
The simple proportion fails.
More advanced models include behavioural effects explicitly.
The act of measurement changed the system being measured.
This is a recurring scientific problem.
Assumption 4: Every Animal Has Equal Capture Probability
Real animals differ.
Males and females move differently.
Adults and juveniles use different habitats.
Dominant animals occupy central territories.
Some individuals are bold.
Others avoid traps.
Animals near roads encounter camera traps more often.
Heterogeneous detection is one of the biggest challenges in abundance estimation.
If a small group of highly detectable animals dominates recaptures, the model may underestimate population because it falsely assumes everyone was equally easy to catch.
Heterogeneity: The Invisible Animals Are Often the Hardest Ones
Imagine 100 animals.
Fifty have detection probability 0.8.
Fifty have detection probability 0.1.
Most observations come from the first group.
A naive model infers that detection is generally high.
It concludes few unseen animals remain.
The low-detection half effectively vanishes.
Modern models use mixture distributions, individual covariates, random effects or spatial capture-recapture to represent heterogeneity.
The most difficult missing data often belongs to the individuals least likely to enter the dataset.
Multiple Capture Occasions: Each Encounter Builds a History
Instead of two samples, run five.
Animal A is seen on occasions 1, 2 and 5.
Capture history:
11001
Animal B appears only on occasion 3:
00100
Each binary sequence contains information about detection and survival.
Summed across hundreds of individuals, capture histories allow much richer models than a single recapture ratio.
Closed-Population Models: Estimate Abundance Over a Short Window
Closed models assume the set of individuals remains fixed during the study.
They focus on abundance and detection probability.
Different model classes allow:
- constant detection;
- time-varying detection;
- behavioural response after first capture;
- individual heterogeneity;
- combinations of these.
Model selection asks which structure the data supports.
The goal is not to make the model as complicated as possible.
It is to represent important sources of detection variation without fitting noise.
Open-Population Models: Let Animals Live, Die, Enter and Leave
Long-term wildlife studies cannot assume closure.
Individuals survive or die.
Young are born.
Animals immigrate and emigrate.
Open-population models such as Cormack–Jolly–Seber and Jolly–Seber families estimate demographic parameters through capture histories.
The question expands from:
How many animals are here?
to:
How many survive, how many enter, and how does the population change through time?
Apparent Survival: Missing Does Not Always Mean Dead
Mark a bird this year.
Do not see it next year.
Did it die?
Maybe.
It may have moved permanently outside the study area.
Many mark-recapture models estimate apparent survival: the probability an individual survives and remains available in the sampled population.
Death and permanent emigration can be statistically confounded without additional information.
The model should say what it can identify.
Not what we wish it could identify.
Temporary Emigration: Sometimes the Animal Leaves and Comes Back
Whales migrate.
Amphibians remain underground.
Large carnivores shift territories.
An animal can be alive and part of the broader population but temporarily unavailable to the sampling method.
Robust-design capture-recapture studies combine short closed sessions within longer open periods.
This separation helps estimate temporary emigration as distinct from mortality.
Study design creates identifiability.
Better data structure can answer questions that more data of the wrong type cannot.
Spatial Capture–Recapture: Where an Animal Was Seen Matters
Traditional capture-recapture treats traps as interchangeable sampling opportunities.
Spatial capture-recapture adds geography.
Each animal has an activity centre.
Detection probability decreases with distance between that centre and a trap or camera.
An animal photographed repeatedly at neighbouring cameras probably lives nearby.
An individual never photographed may simply have an activity centre outside the sampled core.
Spatial models estimate density directly rather than requiring an arbitrary effective sampling area.
Geometry removes one of classical capture-recapture’s hardest ambiguities.
Camera Traps: Capture Without Touching
Motion-triggered cameras can photograph animals continuously with little field disturbance.
For tigers, leopards, jaguars and other individually patterned species, natural markings can identify individuals.
Stripes and spots become marks.
Each photograph records:
- identity;
- camera location;
- time;
- possibly direction and group composition.
The dataset becomes a spatial capture history.
No tranquiliser.
No physical tag.
Mathematics recognises the animal through its coat.
Photo Identification: Pattern Matching Becomes Ecology
Whale flukes have distinctive markings.
Whale sharks have spot patterns.
Zebras have stripes.
Software can extract features and compare photographs against a catalogue.
But misidentification creates statistical error.
If one animal is mistakenly split into two identities, abundance is inflated.
If two animals are merged into one identity, abundance is underestimated.
Identity uncertainty therefore belongs inside the model when error rates are non-negligible.
Genetic Capture–Recapture: A Hair Can Be an Encounter
An animal leaves hair on a rubbing post.
Or faeces on a trail.
DNA genotyping identifies the individual.
Return next week and sample again.
The same genotype appears.
Recapture.
Non-invasive genetic sampling is powerful for elusive carnivores, bears and other difficult species.
But genotyping errors matter.
Allelic dropout can make one individual appear different across samples.
Laboratory quality control becomes part of population estimation.
Acoustic Capture–Recapture: Recognise Individuals by Sound
Some animals have individually distinctive calls.
Microphone arrays can detect and sometimes localise them.
Repeated acoustic detections create encounter histories.
For species that vocalise reliably, sound can substitute for physical capture.
The word capture survives.
The cage disappears.
Distance Sampling: Detection Falls With Distance
Capture-recapture is not the only way to handle imperfect detection.
In line-transect distance sampling, an observer walks a route and records each animal plus its perpendicular distance from the line.
Animals near the line are easier to see.
Farther animals are missed more often.
Fit a detection function describing how probability falls with distance.
Integrate it to estimate effective survey area.
Then infer density.
USGS identifies distance methods, repeated counts and capture-recapture as major families for abundance estimation.
Occupancy Models: Sometimes “Is It Here?” Is Better Than “How Many?”
For rare species, abundance may be too difficult to estimate reliably.
Ecologists may instead estimate site occupancy: the probability a species is present at a site.
Visit the same site several times.
Detection history:
0, 0, 1, 0
The one detection proves presence.
Sites with all zeros are ambiguous.
The species may be absent.
Or present but missed every time.
Repeated surveys separate occupancy from detectability.
Again, absence of observation is not observation of absence.
N-Mixture Models: Repeated Counts Without Individual Marks
Sometimes individual animals cannot be identified.
Researchers repeatedly count at the same site.
Observed counts vary because detection varies even if local abundance remains fixed.
N-mixture models treat true local abundance as a latent variable and counts as imperfect observations of it.
These models are powerful and controversial when assumptions are weak.
Detection and abundance can be statistically difficult to distinguish.
If the model is poorly identified, apparently precise estimates can be misleading.
Complex mathematics does not rescue weak study design automatically.
Integrated Population Models: Combine Several Imperfect Datasets
One survey has capture-recapture data.
Another has nest counts.
A third has harvest records.
Telemetry provides survival information.
Each dataset sees a different part of population dynamics.
Integrated population models combine them in one hierarchical framework.
Breeding data informs recruitment.
Capture histories inform survival.
Counts inform abundance trends.
The whole can estimate parameters more precisely than each dataset alone if assumptions are compatible.
Mathematics lets incomplete datasets reinforce one another.
Maximum Likelihood: Which Parameters Make These Capture Histories Most Plausible?
Suppose survival probability is φ and detection probability is p.
An animal captured in year 1, missed in year 2 and seen in year 3 has a capture history whose probability depends on both survival and detection.
Write the probability of every observed history.
Multiply them to form a likelihood.
Find parameter values that maximise it.
Maximum likelihood converts thousands of encounter sequences into estimates of biological rates.
The animals leave patterns.
The likelihood asks which hidden population process would most plausibly generate them.
Bayesian Models: Preserve Uncertainty Through the Whole Hierarchy
Bayesian capture-recapture models assign probability distributions to unknown parameters and latent states.
Population size is uncertain.
Detection probability is uncertain.
Individual state may be uncertain.
Posterior simulation propagates these uncertainties together.
Instead of one estimate such as 327 animals, the output may say:
median abundance 327, with a 95% credible interval from 250 to 470 under this model.
The interval is not an embarrassing footnote.
It is the honest shape of the knowledge.
Data Augmentation: Add Invisible Possible Animals to the Model
Bayesian abundance models sometimes use a clever trick.
Suppose 120 animals were observed.
Create a “superpopulation” of 500 possible individuals.
The 120 observed animals are certainly real.
The remaining 380 are latent possibilities.
Each receives an indicator variable saying whether it truly belongs to the population.
Inference estimates how many of those invisible rows should be switched on.
The unknown population size becomes the sum of latent membership indicators.
Mathematics gives unseen animals empty rows in a table and lets the data decide how many of those rows are real.
Model Selection: A More Complex Model Is Not Automatically Better
Allow detection to vary by time.
Add sex effects.
Add weather.
Add individual random effects.
Add trap response.
Fit improves.
But complexity also increases variance and overfitting risk.
Information criteria such as AIC compare fit against parameter count.
Bayesian approaches use posterior predictive checking, model comparison and regularising priors.
The goal is explanatory adequacy, not maximum knobs.
Goodness of Fit: Does the Model Reproduce the Patterns It Claims to Explain?
A fitted model can produce a population estimate and still fit poorly.
Check whether simulated data from the model reproduce:
- number of recaptures;
- distribution of encounter frequencies;
- time gaps between sightings;
- spatial distances;
- sex or age patterns.
If the real data contains many individuals seen once and a few seen twenty times, while the model predicts everyone should be seen about equally often, the detection model is wrong.
Residual structure is ecological information.
Population Trend: Sometimes Change Matters More Than Exact Abundance
Conservation often asks whether a population is increasing or decreasing.
If detection probability remains stable, even an imperfect index can reveal trend.
But if detectability changes with habitat, weather or survey technique, the trend can be false.
Long-term monitoring therefore standardises methods or models detection explicitly.
A new camera that sees twice as far can create an apparent population boom.
Measurement change must not be mistaken for ecological change.
Extinction Risk: The Tail of the Distribution Matters
Suppose estimated population is 500 ± 50.
Comfortable.
Suppose it is 70 with a credible interval from 20 to 180.
The lower tail matters enormously.
Population viability analysis combines demographic rates, environmental variation and catastrophes to estimate probability of future decline or extinction.
Conservation decisions depend on tails, not merely averages.
Risk lives where the unlikely and consequential meet.
Harvest Management: How Many Can Be Taken Without Collapsing the Population?
Fishery and wildlife management may regulate harvest.
Set harvest too low and communities lose economic or food benefits.
Set it too high and population declines.
Abundance alone is insufficient.
Managers need recruitment, survival and density dependence.
Adaptive management treats harvest policy as a controlled experiment:
estimate → choose policy → observe response → update model → choose again
The population teaches the management model over time.
Conservation Translocation: How Many Animals Are Enough?
Reintroducing a species into restored habitat raises difficult questions.
Release too few and demographic randomness can eliminate the population.
Release too many and source populations may be harmed.
Sex ratio matters.
Genetic diversity matters.
Survival after release is uncertain.
Population models simulate alternative release strategies.
Mathematics cannot guarantee success.
It can expose which strategies are fragile before animals are moved.
Disease Ecology: Mark–Recapture Can Track Hidden Infection States
An animal may be alive, infected and undetected.
Multi-state capture-recapture models allow individuals to transition among states:
- susceptible;
- infected;
- recovered;
- different geographic locations;
- breeding and non-breeding states.
Some states may themselves be observed imperfectly.
USGS notes that modern capture-recapture frameworks estimate transition probabilities among life-history and geographic states, including uncertain or unobservable states.
Wildlife monitoring becomes a hidden-state model.
Hidden Markov Models: The Animal Has a State We Only Observe Indirectly
A tagged seal can be:
- alive in Area A;
- alive in Area B;
- alive but temporarily unavailable;
- dead.
We do not observe state directly every day.
We observe detections generated probabilistically by state.
Hidden Markov models separate:
- state transition probabilities;
- observation probabilities.
This mathematical architecture appears in speech recognition, finance, genetics and wildlife ecology.
Different worlds.
Same hidden-state problem.
Machine Learning: Identify Animals Faster Without Letting the Classifier Become the Count
Millions of camera-trap photographs create a new bottleneck.
Humans cannot label every image quickly.
Computer vision can classify species and sometimes individuals.
This reduces labour.
But classifier errors propagate into abundance estimates.
If a model systematically misses juveniles because training data contains mostly adults, detection bias becomes biological bias.
Automated labels need validation.
Machine learning speeds observation.
It does not remove sampling theory.
Citizen Science: More Eyes Create More Biases Too
Birdwatchers and nature enthusiasts submit millions of wildlife observations.
The scale is extraordinary.
The sampling is not random.
People visit accessible places.
They prefer charismatic species.
Weekends differ from weekdays.
Observer skill varies.
Statistical models include effort, location, observer and habitat covariates to reduce bias.
More data is valuable.
More data is not automatically representative.
Study Design: Mathematics Begins Before the First Animal Is Seen
Where should traps go?
How many?
How long should each session last?
How far apart should cameras be?
How many recaptures are needed for acceptable precision?
Power analysis simulates expected datasets under plausible abundance and detection rates.
If the planned survey rarely generates enough recaptures to distinguish 100 animals from 300, the study is underpowered before fieldwork begins.
Good statistics saves field effort by identifying weak design early.
Ethics: Better Statistics Can Mean Fewer Captures
Physical capture can stress animals.
Handling carries risk.
Traps can injure.
Researchers therefore minimise handling and use non-invasive methods when appropriate.
More efficient statistical design can reduce the number of animals that need to be captured while preserving inferential quality.
Mathematics improves not only precision.
It can reduce intervention.
A Classroom Thought Experiment: Count the Hidden Beans
Fill an opaque bag with an unknown number of identical beans.
Draw 20.
Mark them with a dot.
Return and mix thoroughly.
Draw 25.
Suppose 5 are marked.
Estimate:
N ≈ 20 × 25 / 5 = 100
Now count the real beans.
Repeat the experiment.
Estimates change.
Plot their distribution.
The child learns sampling variability physically.
Break the Assumption Deliberately
After marking 20 beans, secretly add 30 unmarked beans before the second sample.
Population closure is broken.
The estimator becomes biased upward relative to the original population.
Remove half the marked beans.
Mark loss.
Make marked beans easier to grab by placing them on top.
Unequal capture probability.
Students see why assumptions are not decoration after the formula.
They are the conditions under which the formula means what we claim.
Primary Mathematics: Population Estimation Begins With Fractions
Primary students know:
- fractions;
- ratios;
- proportion;
- sampling;
- graphs;
- averages.
The basic capture-recapture estimator is one proportion.
If 20% of the second sample is marked, perhaps roughly 20% of the whole population was marked.
The sophisticated ecology grows from a Primary-school ratio.
Secondary Mathematics: Probability Enters the Forest
Secondary students can model:
- binomial capture probability;
- confidence intervals;
- bias;
- sampling distributions;
- regression of detection on weather or habitat;
- matrix models of population transitions.
A wildlife survey becomes an applied statistics laboratory.
The forest supplies the hidden variable.
Advanced Mathematics: Ecology as a Hierarchical Inference Problem
Modern population estimation draws on:
- likelihood theory;
- Bayesian inference;
- hidden Markov models;
- spatial point processes;
- Monte Carlo methods;
- matrix population models;
- survival analysis;
- optimisation;
- decision theory.
The biological process is hidden.
The observation process is imperfect.
A hierarchical model separates them.
That separation is the core mathematical achievement.
Why This Improves the World
1. It prevents low visibility from becoming low population
Detection models stop surveys confusing hidden animals with absent animals.
2. It estimates populations without counting every individual
Capture-recapture extracts abundance from partial repeated samples.
3. It measures survival and movement
Long-term encounter histories reveal demographic rates essential for conservation.
4. It makes non-invasive monitoring scientifically powerful
Camera traps, photographs and DNA can substitute for repeated physical handling.
5. It quantifies uncertainty for management
Managers can see whether an abundance estimate is precise or whether the plausible range spans very different conservation decisions.
6. It turns monitoring into adaptive learning
New observations update survival, abundance and policy models over time.
What Mathematics Does Not Do
Capture-recapture does not make every assumption true.
It does not guarantee equal detectability.
It does not distinguish death from permanent emigration without enough information.
It does not repair biased sampling design.
It does not turn a camera classifier into a perfect individual identifier.
It does not decide how much habitat should be protected or how human livelihoods should be balanced with conservation.
And a precise population estimate is not the same as a healthy ecosystem.
Frequently Asked Questions
What is capture–recapture?
Capture–recapture is a family of methods that repeatedly samples identifiable individuals. Recapture frequencies are used to estimate population abundance, detection probability, survival, movement or other demographic parameters.
Why can’t scientists just count the animals they see?
Because detection is imperfect and varies among habitats, individuals, observers and conditions. Raw counts combine true abundance with detectability and can therefore misrepresent population size or trends.
Does an animal have to be physically captured?
No. Capture can mean any reliable encounter with an identifiable individual, including camera-trap photographs, natural markings, acoustic identification or DNA genotypes.
What is spatial capture–recapture?
Spatial capture–recapture uses both individual encounter histories and the locations of traps or detections. Detection probability is modelled as decreasing with distance from an individual’s latent activity centre, allowing population density to be estimated explicitly.
Why are repeated surveys important?
Repeated observations separate presence from detection and help estimate survival, movement and heterogeneity. One survey cannot reveal whether a missing animal was absent or simply missed.
Sources and Further Reading
- U.S. Geological Survey, Capture-Recapture Methodology, an overview of abundance, survival, transition models and modern non-invasive marking approaches.
- U.S. Geological Survey, Capture-Mark-Recapture Science, explaining marked-to-unmarked ratios and applications to population dynamics.
- U.S. Geological Survey, Abundance: Population Size and Density Estimation, reviewing distance, repeated count and capture-recapture approaches and the central role of detection probability.
- Otis, Burnham, White and Anderson, Statistical Inference from Capture Data on Closed Animal Populations, foundational work on heterogeneous capture probabilities and closed-population models.
Continue Through eduKateSG
Continue with How Mathematics Works. Compare this article with Finding a Planet We Cannot See: both infer hidden populations or objects from incomplete detection. Also compare it with Seeing Inside Without Cutting Open, where observations are again separated from the hidden state that generated them.
Final Thought: The Animal You Did Not See Is Still Part of the Science
A tiger walks past no camera.
No photograph exists.
A whale is alive but does not surface near the survey vessel.
A bird survives the winter and moves beyond the study boundary.
The missing observation is not nothing.
It is part of a probability.
And once detection probability becomes explicit, a conservation scientist can ask a far better question than “How many did we see?”
How many were likely there?
How certain are we?
What changed?
And what should we do before the next survey?
Mathematics improves the world here by refusing to let invisible animals disappear merely because the observer failed to see them.