Industry: Environmental Product: ModelRisk Application: Habitat preservation under uncertainty
A conservation trust manages a reserve built as a chain of six habitat patches — a protected nucleus joined by five corridors to progressively smaller satellite patches. The number that decides whether the focal species persists is not the total land on the title deed; it is the effective connected habitat area: the usable core of every patch reachable from the nucleus through corridors that are actually intact, after edge effects have eaten into each patch's interior. The trust's planning map adds up the patch cores assuming every corridor is open and edge loss is average, and reports 1,414 connected hectares — almost three times the 500-hectare minimum viable area (MVA) the species needs. Connectivity looks like a solved problem.
The map assumes a benign year. Corridors do not stay open by decree: a drought or fire season raises every corridor's failure probability and deepens edge loss across every patch simultaneously. Simulate the reserve over 60,000 landscape years in ModelRisk and the static 1,414-hectare figure becomes a wide distribution with a dangerous floor. The mean connected area is 754 ha, the median 677 ha, the P10 just 264 ha — and the effective connected area falls below the 500-ha MVA in 33.9% of years. The reserve that looks twice as large as it needs to be is below viability roughly one year in three.
Connectivity is a property of the worst years, when corridors break and edges retreat together — and a deterministic map, drawn for an average year with all corridors open, reports a number that essentially never occurs. Three uncertainties drive the gap, and a point estimate hides all three.
The structural one is the shared landscape disturbance factor, and it is also the modelling trap. If each corridor failed independently and each patch's edge loss were its own private draw, the failures would average out and the connected-area spread would be tame. They are not independent: a single drought or fire severity — one Beta-distributed draw per landscape-year, mean about 0.36 — simultaneously raises every corridor's failure probability and deepens every patch's edge loss. Because corridors fail together in a bad year, the realised pairwise correlation between corridor failures is +0.102, and the joint exceedance of disturbance and breakage is far heavier than independence predicts. The other two uncertainties are patch area, lognormal with shape 0.35 around each patch's median, and edge-effect loss, a Beta-distributed fraction of each patch eaten by the perimeter, both of which the planning map collapsed to single values.
The reason the lower tail is fat is that the two things that destroy connectivity are not independent — they are driven by the same landscape year. Plotting disturbance severity against the number of broken corridors makes the dependence visible: the probability of landing in the corner where the year is harsh AND two or more corridors break is 23.5%, against the 17.1% a naive independence assumption would predict. That extra weight — roughly a third more than independence — is precisely where the sub-MVA years cluster. A model that multiplies marginal corridor-failure probabilities together would miss it entirely.
The decision is where to spend a limited preservation budget: harden the existing corridors (fencing, underpasses, fuel breaks that lower link failure), or acquire and enlarge land. ModelRisk answers it as a curve — P(connected area meets the MVA) against investment intensity.
From the base probability of 0.66 above the MVA, corridor hardening alone tops out at 0.70 — it cannot reach a 90% viability target on its own, because no amount of link reliability protects against a harsh year that also strips patch cores through edge loss. Land acquisition that enlarges the patch areas does reach 90%, but only at roughly double the current patch area. The curve reframes the budget conversation: the reserve cannot be made reliably viable by maintenance spending; it needs more core area, and the simulation sizes how much.
Swinging each input across a plausible range and measuring the effect on P(connected area below MVA) ranks the levers. Landscape disturbance severity dominates at a 26.9 percentage-point spread (from 17.3% to 44.2% sub-MVA between a mild and a harsh disturbance regime), followed by edge-effect loss at 16.9 points and corridor reliability at 10.1 points. Patch-area uncertainty is nearly inert at 0.8 points. The ranking says the dominant risk is the landscape regime itself — which argues for fire and drought resilience (fuel management, water buffering) and for the extra core area that lets the reserve absorb a bad year, rather than for incremental corridor tinkering.
A reserve's connectivity is not the sum on the planning map — it is a distribution with a fat low tail driven by the years when everything breaks at once, and preservation has to be designed against that tail, not against the benign-year total.