Metapopulation ecology studies how local populations of the same species interact through dispersal, and how their combined dynamics affect persistence across a landscape. A population can disappear from one habitat patch while the species survives elsewhere and later recolonises that patch. Regional survival therefore depends on more than the number of individuals counted at one place: it also depends on where suitable habitat remains, which populations supply dispersers and whether arrivals establish successfully.
The gold standard is to connect those mechanisms to evidence at the right spatial and temporal scales. A map of occupied patches is a starting point, not a demonstration that a network is secure. This guide explains patch turnover, regional thresholds, source–sink relationships, network structure, observation error and changing environments. All numerical examples are original fictional teaching cases with explicit assumptions; they are not forecasts for a particular species or landscape.
Begin with the Gold Standard master for the wider collection, the Education guide for learning that transfers, or the How X Works hub for other mechanisms. The Ecology root supplies the broader relationships among organisms, populations and environments.
Choose your metapopulation ecology route
Follow local populations through patch turnover, regional persistence, network structure, survey evidence and changing conditions.
Route 1 · Chapters 1–3: Define populations and patch turnover
Route 2 · Chapters 4–6: Calculate persistence and network effects
Route 3 · Chapters 7–9: Connect movement, surveys and evidence
Route 4 · Chapters 10–11: Follow environmental change and decisions
Route 5 · Chapters 12–14: Transfer, diagnose and continue
Chapter index
Define populations and patch turnover · Chapters 1–3
Calculate persistence and network effects · Chapters 4–6
Connect movement, surveys and evidence · Chapters 7–9
Follow environmental change and decisions · Chapters 10–11
Transfer, diagnose and continue · Chapters 12–14
A metapopulation is not simply a species shown in several coloured areas on a map. The areas must correspond to local populations whose dynamics are meaningfully connected by dispersal. A useful patch boundary separates places where births, deaths and routine movements form a local demographic process from movements that connect that process to another population. What counts as local depends on the organism, its life cycle and the question.
A small meadow might contain a breeding population of an insect with limited routine movement. The same meadow might be only a feeding stop for a bird that breeds across a much larger area. Counting the bird's visits as repeated colonisations would misdescribe its biology. Conversely, two visually continuous areas could be demographically separated by a barrier that a focal organism rarely crosses.
Write an operational definition
For a fictional meadow insect, define a patch as a mapped area containing its larval food plant and suitable breeding conditions. Define occupancy as evidence of a locally established breeding population during a specified annual survey period. Record the map version and treatment of patches that merge, split or lose their food plants. This makes an apparent disappearance interpretable: was the population lost, the habitat destroyed, or the survey unit changed?
The classic extinction–colonisation picture is especially informative when local populations sometimes disappear and empty suitable patches can be recolonised. Other spatial population systems have a persistent large population feeding smaller ones, or such frequent exchange that a nearly continuous population model is more appropriate. These are reasons to choose a different model, not reasons to force the observations into a familiar diagram.
Metapopulation ecology usually follows one focal species. Metacommunity ecology follows multiple interacting species across connected local communities. For the latter question, use the specialist guide to metacommunity dynamics, dispersal limitation and mass effects. Keeping the unit explicit prevents a result about occupancy of one species from being mistaken for a result about total biodiversity.
Dispersal connects local populations, but movement is only one part of the process. An individual must leave, survive the journey, reach a useful destination and contribute to the relevant population outcome. A photograph of an arriving adult establishes arrival. It does not by itself establish breeding, recruitment or a self-sustaining population.
Colonisation is establishment in a previously unoccupied patch under the chosen definition. Recolonisation follows an earlier local loss. A rescue effect occurs when immigration reduces the chance that an existing local population disappears. These processes differ in their starting state: colonisation begins with an empty patch, whereas rescue acts on a population that is still present. Brown and Kodric-Brown formalised the effect of immigration on extinction in their 1977 rescue-effect paper.
A demographic budget exposes the distinction
Suppose a patch starts with 40 individuals. During a defined interval there are 12 births, 18 deaths, eight immigrants and two emigrants. The ending count is 40 + 12 − 18 + 8 − 2 = 40. The stable count conceals a local birth–death deficit of six, offset by net immigration of six. Counting only the beginning and ending populations would miss that dependence.
This is a single-interval budget, not proof that the patch is a persistent demographic sink. Rates can vary with density, age structure, weather and year. Establishing source or sink status needs the relevant local demographic performance and its dependence on immigration over an appropriate period. Pulliam's 1988 source–sink model explains how productive habitats can maintain populations where local reproduction does not replace mortality.
The diagnostic question is therefore precise: did immigrants establish a new population, prevent an existing population's loss, or maintain a local demographic deficit? Different answers imply different measurements and different consequences if the supplying population declines. A busy patch can depend strongly on another patch that appears less conspicuous to a casual observer.
Imagine 40 suitable patches surveyed with perfect detection at the beginning and end of a year. Twenty-four are initially occupied. Six of those lose their populations. Five of the 16 initially empty patches become occupied. The final total is 24 − 6 + 5 = 23 occupied patches, so occupancy falls from 60% to 57.5%.
The local extinction proportion is six divided by 24, or 25%. The colonisation proportion is five divided by 16, or 31.25%. Dividing both counts by 40 would instead give the fraction of all patches undergoing each transition. Those are legitimate descriptive quantities, but they answer different questions. The denominator identifies the state from which a transition was possible.
Small net change can conceal substantial reorganisation
Only one fewer patch is occupied at the end, but 11 patches changed state. The regional count alone therefore misses how much the occupied set changed. If the six losses included important disperser-producing patches and the five gains were isolated, the next year's prospects might worsen even though total occupancy changed little. The reverse arrangement could have different consequences.
This two-survey example also hides events within the interval. A patch could lose and regain its population between visits and appear continuously occupied. Shorten the interval and more events may become visible, but survey effort and disturbance can increase. The appropriate interval follows the biological process, not an arbitrary preference for monthly or annual data.
Keep three outputs separate: the fraction of patches occupied, the number of state changes and the number of individuals within occupied patches. A region could retain the same occupancy while abundance collapses within most patches. For a slower introduction to the occupancy accounting, the existing How Habitat Patches Stay Connected Through Local Populations tutorial develops the patch-transition logic step by step.
A simple continuous-time model represents occupancy by p, the fraction of suitable patches occupied. Let c be a colonisation coefficient and e a local extinction rate, both measured per unit time. The model is dp/dt = cp(1 − p) − ep. Occupied patches supply colonists, the fraction 1 − p offers empty destinations, and occupied patches experience losses. This is the standard simple occupancy formulation associated with Levins's 1969 work.
For a positive equilibrium, divide the zero-change equation by p: c(1 − p) − e = 0. Rearranging gives p* = 1 − e/c. A positive equilibrium requires c greater than e. Zero occupancy is also an equilibrium because the model contains no external source of colonists.
Work an original parameter example
Choose c = 0.60 per year and e = 0.15 per year. The positive equilibrium is 1 − 0.15/0.60 = 0.75. At p = 0.40, colonisation contributes 0.60 × 0.40 × 0.60 = 0.144 per year and extinction removes 0.15 × 0.40 = 0.060 per year. The instantaneous net change is +0.084 per year: 8.4 percentage points of occupancy per year at that state.
That derivative is not an exact prediction that occupancy will be 48.4% one year later. The rate changes as p changes. Likewise, c is not the annual probability that any given empty patch is colonised: colonisation pressure in this model depends on p. Continuous rates and finite-interval probabilities cannot be interchanged without specifying the transition process.
The model deliberately compresses patch differences and spatial arrangement. It assumes fixed suitable habitat, homogeneous rates and a mixing approximation in which regional occupancy summarises colonist supply. It has no explicit population sizes, weather shocks or observation error. The threshold is a conditional mathematical result. Its value is that it makes a mechanism testable; it does not certify the safety of a real landscape from two estimated numbers.
CHAPTER 5 OF 14
5. Distinguish deterministic persistence from finite-population risk
The positive equilibrium in the previous chapter describes a deterministic model. A real network contains a finite number of patches and individuals. Random birth, death, dispersal and environmental events can remove all remaining populations. If the network is closed and all patches become empty, recolonisation cannot begin internally. The all-empty state is absorbing unless an outside source or an explicitly modelled dormant stage can restore the population.
A three-patch risk calculation
Suppose three patches are currently occupied. In a deliberately simplified annual model, each has a 0.20 probability of losing its population, losses are independent, and there is no recolonisation during the interval. The probability that all three disappear is 0.20 × 0.20 × 0.20 = 0.008, or 0.8%. A low one-year probability is still not zero, and it says nothing by itself about a century of changing conditions.
Now keep the same marginal 20% loss probability for each patch but make loss perfectly synchronised: a shared event removes all three populations with probability 0.20, otherwise none disappears. Regional loss is now 20%. The difference is dependence among patches, not a change in the individual patch probabilities. This constructed contrast explains why three populations exposed to the same failure mechanism do not provide three independent safeguards.
Do not calculate a many-year risk by blindly repeating the initial one-year value. After one patch disappears, the state changes; colonisation may occur, habitat may deteriorate, and the weather process may have memory. A proper finite-horizon calculation follows possible states and transitions through time, or simulates an explicitly specified demographic model.
This is where population viability analysis, extinction risk and model uncertainty becomes useful. State the event of concern, the time horizon, the initial state and the model assumptions. Probability of complete regional extinction, probability of falling below five occupied patches and expected occupancy after ten years are different outputs and should not share the same label of viability.
Two landscapes can contain the same total habitat area but arrange it differently. A focal species may reach nearby patches readily and rarely cross a wide gap. Patch size can also influence local population size or extinction risk, while quality affects breeding success. Area, arrangement and biological performance therefore need explicit roles rather than being combined into an unexplained score.
Hanski and Ovaskainen's 2000 metapopulation-capacity study developed a model-based way to combine patch areas and connections. Its capacity measure is the leading eigenvalue of an appropriate landscape matrix, compared with a species-dependent threshold. It is not a universal map index that retains the same biological interpretation when its assumptions or matrix definition change.
A small network calculation without hidden software
Consider a separate teaching model with three identical patches. Near low occupancy, each occupied patch contributes colonisation pressure b to each connected neighbour, while every patch has loss rate e. Write the linear approximation as dx/dt = (bA − eI)x, where x contains patch-occupancy probabilities, A records connections and I is the identity matrix. This is an illustration of network growth, not a fitted application of the published capacity model.
In a triangle, every patch has two neighbours. For the equal vector (1,1,1), multiplication by A gives (2,2,2), so the leading eigenvalue is 2. In a three-patch chain, the leading eigenvalue is the square root of two, approximately 1.414. It can be checked directly: A acting on (1,√2,1) gives (√2,2,√2), which is √2 times the original vector.
Set b = 0.10 per year and e = 0.16 per year. Low-occupancy growth is 0.10 × 2 − 0.16 = +0.040 per year for the triangle, but 0.10 × 1.414 − 0.16 ≈ −0.0186 per year for the chain. Identical patch counts give opposite model predictions because their connections differ. Real networks require weighted, often directional connections and a biological account of their costs; the diagram alone does not supply those values.
A straight-line distance is easy to calculate. Successful dispersal is harder to establish. Two patches separated by 500 metres of usable cover may exchange more individuals than two separated by 200 metres of hostile terrain. Wind, currents, roads, vegetation, behaviour and life stage can make movement asymmetric. A connection from A to B need not imply the same connection from B to A.
The Landscape Ecology root distinguishes mapped structure from functional connectivity. For metapopulation work, take that distinction one step further: functional movement must connect to colonisation, rescue or reproduction if those are the outcomes being predicted. Tracking an individual across a corridor demonstrates movement; documenting successful recruits addresses a later stage of the mechanism.
A transparent distance-weight example
For an illustrative connection weight, choose w = exp(−d/L), where d is distance and L is a characteristic distance scale. This is a hypothesised decay rule, not a measured dispersal probability. With L = 1 kilometre, a 0.5-kilometre separation gives exp(−0.5) ≈ 0.607, and a 2-kilometre separation gives exp(−2) ≈ 0.135. The second weight is about 22% of the first.
If the organism's effective L were only 0.2 kilometres, the same 0.5-kilometre gap would receive exp(−2.5) ≈ 0.082. The physical map did not change; the biological interpretation did. This is why a landscape judged connected for one species cannot automatically be judged connected for another.
A stepping stone is useful only if it serves the needed process. It might support breeding, permit rest during transit or simply reduce exposure, and these functions imply different models. An apparently central patch may also have poor habitat quality. Evaluate the whole causal chain before treating a proposed link as a substitute for maintaining productive habitat at its ends.
Not detecting an organism is not equivalent to confirming that its population is absent. Survey effort, observer skill, weather, life stage and abundance can affect detection. If detectability drops after a habitat change, an apparent rise in local extinction can partly reflect observation rather than demography. If detectability improves, apparent colonisation can include populations that were present all along.
MacKenzie and colleagues' 2003 dynamic occupancy study estimates occupancy transitions while allowing imperfect detection, using repeat surveys within primary periods. The central distinction is between an unobserved ecological state and observations conditional on that state. The site's dynamic occupancy models guide develops the survey assumptions in more detail.
Calculate what three missed detections mean
Assume a patch is occupied throughout three visits, detection probability is 0.60 on each visit, and detections are conditionally independent. The chance of missing the population three times is (1 − 0.60)³ = 0.064. Thus the probability of at least one detection is 93.6%. Neither number is the probability that a patch is occupied after observing three zeros.
For that latter question, add a prior occupancy probability of 0.70 and assume no false positives. The probability of three zeros is 0.70 × 0.064 + 0.30 × 1 = 0.3448. The probability of occupancy given three zeros is 0.0448/0.3448 ≈ 0.130, or 13.0%. Changing the prior probability or detection process changes the answer.
These are elementary conditional-probability calculations, not a substitute for fitting a survey model. If all three visits share an unsuitable weather condition, independence may fail. If colonisation or extinction occurs between visits that were assumed to share a state, closure fails. A good diagnosis first checks whether the survey design could distinguish the proposed ecological transitions from the ways observations were generated.
CHAPTER 9 OF 14
9. Use long-term evidence without turning one system into a universal rule
The Glanville fritillary butterfly in Finland's Åland Islands provides a major empirical metapopulation study system. Ojanen and colleagues' 2013 account of its monitoring programme describes repeated surveys of thousands of meadow patches and the practical demands of maintaining long-term records. The strength of such work comes from linking mapped habitat, repeated population observations and additional biological measurements over time.
A 2017 study by Hanski and colleagues analysed 22 years of observations across 4,415 small dry meadows. It examined persistence in heterogeneous patch networks and relationships involving habitat configuration, quality and genetic variation. The appropriate lesson is that regional dynamics can depend on several connected mechanisms, not that patch area or a single genetic marker gives a universal prediction for other species.
Ask what each kind of evidence can establish
A repeated occupancy record identifies a pattern of observations. A demographic study can help explain local production and loss. Movement records can inform connections. Genetic data can reveal another aspect of population relationships, but historical gene flow need not equal present-day demographic rescue. Combining evidence is useful when the time scales and mechanisms align.
Suppose a model predicts low colonisation for isolated patches and the observations agree. That agreement is informative, but isolation could also correlate with poor habitat or low survey effort. A stronger analysis measures plausible alternatives, tests predictions beyond the data used to fit the model, and checks whether its conclusions depend on a few influential patches.
For a reader assessing a study, record the focal species, patch definition, observation period, transition model, treatment of detection and validation strategy. Then state what would make the conclusion fail in a new landscape. This exercise is more useful than memorising that butterflies are a classic example: it teaches how a field system supports, constrains and sometimes changes a general explanation.
A patch network is not necessarily fixed. Succession can remove breeding habitat, a disturbance can create it, and management can alter its quality. Newly suitable habitat may remain empty because colonists have not arrived. A population can remain present for a time after conditions deteriorate, so current occupancy can lag behind the processes governing its future.
Such delays matter for interpretation. A declining network does not instantly become empty at the moment a model threshold is crossed. Conversely, restoring habitat does not guarantee immediate recovery. Colonist supply, establishment and time to reproduction may all delay a response. A monitoring period that ends too early can confuse a slow response with no response, or temporary persistence with long-term security.
Shared conditions can erode the benefit of multiple patches
Tack, Mononen and Hanski's 2015 study of low summer precipitation analysed long-term Glanville fritillary dynamics and linked drought during a vulnerable life-history stage with increasing spatial synchrony and fluctuation amplitude. This is evidence from a particular system, not a claim that every dispersing species responds identically to drought.
An original planning thought experiment makes the general issue visible. Four nearby wetlands may all depend on the same water supply. Their separate map outlines do not make their drying risks independent. A more distant wetland with a different water regime could offer environmental diversity, but its distance might also reduce recolonisation. Evaluating the tradeoff needs both environmental dependence and movement information.
Avoid replacing a variable process with an average merely because the average is easy to report. Two landscapes can have the same average annual occupancy but different frequencies of very low occupancy, different recovery times and different dependence on a few surviving patches. Report the feature that answers the decision: maintaining a typical population size and avoiding a rare regional collapse are related but distinct goals.
A useful decision question is concrete: which feasible action is expected to improve a defined persistence outcome over a stated period, compared with what would happen without it? Enlarging a productive patch, improving quality, retaining an empty stepping stone and reducing a movement barrier act through different mechanisms. Their costs, delays and uncertainties also differ.
A deliberately simple comparison
Return to the model dp/dt = cp(1 − p) − ep with c = 0.60 and e = 0.15 per year. Its equilibrium occupancy is 0.75. Suppose a habitat-quality action is expected to reduce e to 0.12 while leaving c unchanged. The new equilibrium is 1 − 0.12/0.60 = 0.80. A different connectivity action that raises c to 0.75 while leaving e unchanged also gives 1 − 0.15/0.75 = 0.80.
Equal predicted equilibrium occupancy does not make these actions equivalent. Starting at p = 0.40, the quality action gives a derivative of 0.60 × 0.40 × 0.60 − 0.12 × 0.40 = 0.096 per year. The connectivity action gives 0.75 × 0.40 × 0.60 − 0.15 × 0.40 = 0.120 per year. Their early trajectories differ, even in this simplified model.
The assumed parameter changes would need evidence. A corridor might not change establishment, and a quality intervention might affect both disperser supply and extinction. A real comparison should test uncertain effects, implementation delays and outcomes that the occupancy model omits. A small difference between predicted actions may be less informative than the uncertainty around either prediction.
For evaluation, compare suitable treated and untreated networks over time where the design permits it. Record changes in habitat and detection as well as population outcomes. Nearby controls may receive dispersers from treated patches, so the intervention can affect the comparison itself. The Conservation Biology root connects these ecological mechanisms to intervention evidence, monitoring and practical responsibilities.
CHAPTER 12 OF 14
12. Transfer check: the same map under a different dispersal regime
Use the three-patch network from Chapter 6. Initially, a triangular network has b = 0.10 and e = 0.16 per year. Its low-occupancy growth rate is +0.040 per year. Now suppose the focal organism changes, or credible new evidence shows much lower successful movement. The map remains identical, but b is 0.06 per year. Assume the same loss rate solely to isolate the effect of movement.
Before reading the answer, decide whether the original persistence argument survives. Then compare two hypothetical responses: restoring b to 0.10, or reducing e to 0.10 while leaving b at 0.06. Finally, repeat the quality response for a chain rather than a triangle.
Answers and interpretation
With reduced movement, the triangle's growth rate is 0.06 × 2 − 0.16 = −0.040 per year. The old conclusion fails despite an unchanged map. Restoring b to 0.10 restores +0.040 per year. Reducing e to 0.10 instead gives 0.06 × 2 − 0.10 = +0.020 per year. Both changes cross this model's low-occupancy threshold, but their predicted rates differ.
For the chain, reducing e to 0.10 gives 0.06 × √2 − 0.10 ≈ −0.0151 per year. The same quality change is insufficient in that network under these assumptions. This is a changed-context test: the calculation must be rebuilt from the relationship between movement, losses and network structure, rather than carried over from a memorised answer.
The limits travel with the calculation. Positive low-occupancy growth in the deterministic approximation is not a quantified finite-network survival probability. Nor does it demonstrate that a proposed intervention can achieve the parameter change. A defensible conclusion names the model result, identifies the missing empirical evidence and avoids converting an illustrative threshold into a management guarantee.
Does an empty patch have no conservation value?
No. It may be suitable but temporarily unoccupied, a future destination after local loss elsewhere, or part of a changing habitat network. It may also be unsuitable or effectively unreachable. Determine which explanation fits the evidence before assigning its value. An empty point on one survey map is not enough to choose among these possibilities.
Does high occupancy prove regional security?
No. Occupied populations could be small, dependent on immigration, exposed to the same disturbance or declining after habitat deterioration. Check abundance where relevant, local demographic performance, turnover, environmental dependence and the time horizon. Occupancy is a useful state variable when its relationship to the question is made explicit.
Are more connections always better?
The answer depends on what moves and what limits persistence. Connections can enable recolonisation and rescue, but they can also alter emigration losses and connect populations to shared disturbances or other organisms. The simplified positive connection weights in the teaching model leave such processes out. Their omission is a model choice, not evidence that they are absent in nature.
Is every source permanently good and every sink permanently bad?
No. Demographic performance can change with density, year and life stage. A sink population can still contribute to wider ecological relationships, and a source classification does not by itself determine a conservation decision. Diagnose the local birth–death balance and movement dependence, then evaluate the stated regional goal.
What should be checked when a model fits occupancy but misses turnover?
Inspect colonisation and extinction separately. Excess losses and excess gains can cancel in the net total. Check whether patch quality, connectivity and detection vary, whether within-interval transitions are missed, and whether the model predicts the identities of changing patches as well as the regional count. Agreement with one summary does not establish agreement with the process.
What is the quickest repair for a confused calculation?
Write the state, denominator, time unit and assumption beside each number. Label a rate as a rate, a probability as a probability and a count as a count. Recompute a simple boundary case: when no patches are occupied in a closed model, internal colonisation must be zero. When a formula violates that boundary, either it includes an external source that needs explaining or its interpretation is wrong.
Begin with a paper landscape of six patches and a fictional annual record. Mark the patches occupied in each year, count losses only among those initially occupied and count gains only among those initially empty. Then remove one survey visit or lower detection probability and explain what can no longer be inferred. The goal is a correct relationship between process and evidence, not a complicated map.
Next, reproduce the threshold and network calculations in a spreadsheet. Keep inputs separate from formulas and label every time unit. Vary c, e or b one at a time, then vary plausible combinations. A useful output states where the conclusion changes and which uncertain measurement most affects that boundary. It should also show when the model cannot answer the question at all.
For a field-reading exercise, choose a published study and reconstruct its reasoning without collecting or moving organisms. Identify the patch definition, population state, survey method and evidence of connection. Distinguish data supporting local persistence from data supporting regional persistence. This makes the next learning step concrete while respecting the limits of an introductory analysis.
Primary sources and what each contributes
Levins, 1969 provides an early mathematical treatment of population persistence across heterogeneous environments. Brown and Kodric-Brown, 1977 explains the rescue effect. Pulliam, 1988 develops source–sink population regulation. Read them for their mechanisms and assumptions rather than as ready-made prescriptions.
Hanski and Ovaskainen, 2000 connects spatial patch networks to metapopulation capacity. MacKenzie and colleagues, 2003 addresses estimation with imperfect detection. Ojanen and colleagues, 2013, Tack and colleagues, 2015 and Hanski and colleagues, 2017 connect long-term field records with spatial, environmental and biological explanations.
Continue through Landscape Ecology to examine habitat maps, dynamic occupancy models to examine observations, or population viability analysis to examine finite-horizon risk. The practical standard remains the same: define the population system, follow the mechanism, quantify the uncertainty and check whether the conclusion survives when the context changes.