Bird surveys, collision models and the cost of getting biodiversity risks wrong

In Part 1, we looked at why traditional bird surveys are stalling wind development in the UK and Ireland, and why NatureScot's recommended minimum of 72 per vantage point a year covers under 2% of the year's daylight and none of its darkness.

That small window of observation becomes the basis for a collision risk model which influences decisions, mitigation requirements and ultimately millions of pounds of project exposure. Consultancies follow the methodology. Regulators trust the outputs. Developers sign off on mitigation packages built on them.

But once projects become operational, existing monitoring methods rarely provide enough evidence to establish whether those initial predictions matched reality.

What happens when predictions go unverified

Between October 2024 and May 2025, three white-tailed eagles were killed by turbines at wind farms in south Donegal. Two males were struck by the same turbine. These are birds from Ireland's reintroduction programme, which has released 245 chicks since 2007 to bring the species back from national extinction. Every one of these projects passed through consenting, with surveys and collision risk assessments that did not foresee this outcome.

And we only know about any of it because the eagles happened to carry satellite tags. The Department of Housing, Local Government and Heritage acknowledged that mortality may be higher, because "the deaths of untagged eagles may go unrecorded if removed by scavengers."

Hornsea 3 in the North Sea was consented only after Ørsted committed to a kittiwake compensation package reported at around USD 20 million: artificial nesting structures, quickly nicknamed "kittiwake hotels," built to offset collision mortality that a model predicted might happen offshore. A 2.85 GW project and a multi-million-pound compensation scheme both hinged on modelled collision numbers that no one had ever validated against observed mortality. Kittiwakes are red-listed in the UK and deserve serious protection. But if predicted impacts are never validated once the project is operational, how confidently can an impact assessment justify the scale of compensation and consent requirements attached to them?

Two projects, two opposite errors, same root cause: a small amount of observed data is being asked to carry a large amount of modelled assumption.

Why are collision risk models so far off?

A collision risk model is only as good as the flight activity data fed into it. In the UK and Ireland, that data largely comes from vantage point surveys following NatureScot guidance, whose limitations we looked at in Part 1. The guidance recommends a minimum of 36 hours per vantage point in each of the breeding and non-breeding seasons, and more where a site is particularly sensitive. Against roughly 4,400 hours of annual daylight, 72 hours per vantage point is under 2%. No nights, no fog, no high winds, no observer fatigue corrections. The remaining 98% is assumption.

The Band model then multiplies that thin sample by avoidance rates, generic correction factors that swing the output by orders of magnitude. Small changes in assumed avoidance produce wildly different predicted mortality rates.

At Aberdeen Bay, 19 months of continuous AI-based camera monitoring recorded over 137,000 bird detections and 2,007 tracked flights near a turbine, running through roughly 95% of daylight hours. The pre-construction assessment had predicted between 2.18 and 8.54 collisions per turbine per year. The observed count over 19 months: zero. The original prediction was off by more than three orders of magnitude.

The same framework, fed real data instead of a 2% sample, produced an answer thousands of times smaller. And in Donegal, the same class of methodology produced answers that were fatally too small. The models are not consistently cautious or consistently lax. They are consistently uncalibrated, in both directions.

Post-construction monitoring does not give ground truth either

The uncertainty does not disappear once a project is operational. Post-construction monitoring still relies heavily on sampled carcass searches that leave substantial blind spots.

Braes of Doune Wind Farm in Scotland ran one of the UK's more intensive red kite monitoring programmes: 1,486 hours of vantage point observation, radio tagging, turbine searches and carcass-persistence trials, with searcher-efficiency trials finding 81% of test carcasses. Three red kites were documented as turbine-strike fatalities. Only one was found by the systematic turbine searches.

That was not a poorly designed monitoring programme. Quite the opposite. It demonstrates why the absence of a carcass cannot simply be interpreted as the absence of a collision. Searcher efficiency, scavenging, vegetation, search area, frequency and chance all sit between mortality occurring and mortality being recorded.

If an intensive programme needs correction factors to reconstruct what probably happened, routine operational monitoring deserves more scepticism. Donegal shows what that means in practice: without satellite tags, three dead eagles would likely never have entered any dataset.

The developer owns the uncertainty

Once the consultancy completes the EIA work and consent is granted, the developer lives with the result for decades.

If risk has been underestimated, that can mean unexpected collisions, regulatory scrutiny, additional monitoring and mitigation measures. If it has been overestimated, developers can accept conservative design decisions, compensation requirements and operating constraints against an impact that may later prove materially different from reality.

And when nobody is sufficiently confident either way, projects can be sent back for another survey season or another layer of assessment.

Hornsea 3's artificial nesting structures are not evidence that the compensation was unnecessary. They are evidence of the financial scale that environmental modelling and uncertainty can eventually reach. The BTO later found out that compensation calculations across developments were inconsistent, highly sensitive to demographic assumptions and constrained by limited empirical data. It also found that the evidence for whether these structures work as compensation at all remains limited. Those findings deserve more attention and should call into question the confidence placed in current assessment methodology to justify the compensation requirements imposed on projects.

We are asking relatively small datasets to support increasingly large decisions. Eventually, that mismatch becomes somebody's problem. Usually the developer's.

We have changed "standard methodology" before

Offshore ornithology has already been through a similar transition. There was a time when offshore bird surveys meant observers on boats scanning the sea and sky through binoculars. Digital aerial surveys arrived with a very different proposition: collect imagery first, then allow ornithologists to identify, verify and revisit the evidence afterwards. East Anglia ONE became the first offshore wind farm to gain planning approval relying solely on digital aerial bird surveys for its EIA.

Today, NatureScot describes digital aerial surveys as the preferred method for collecting the relevant offshore bird-distribution data. Regulators continue to scrutinise how the technology, survey design and analysis should evolve. The ornithologists did not become irrelevant. The evidence changed. AI, cameras and other automated monitoring technologies are at an earlier point on that same curve today, and still need testing. But waiting for a new method to become written into guidance before questioning the limitations of the current one risks putting the process before the evidence.

NatureScot's own guidance already recognises that its methods are not exhaustive: it provides a standardised approach to collecting evidence, not a guarantee that the evidence will capture everything that matters at a particular site.

That puts environmental consultancies in an important position. They are the people closest to the data. They see the hours that were observed and the thousands that were not, and they know when a flight-activity estimate rests on a handful of observations. They see where assumptions enter a CRM and how those assumptions carry through into conclusions that developers and regulators are expected to rely on.

So perhaps the question should not simply be: have we collected enough data to comply with the methodology? It should be: do we have enough data to stand behind the risk we are being asked to predict? The methodology may define what has to be measured, but cannot decide when we know enough.

Spoor works with consultancies and developers on continuous, camera-based bird monitoring. If you have a site where the evidence feels thin, get in touch.