How Machine-Learning-Derived Land Cover Differs From Surveyed Data
Overview
Two habitat maps can show the same polygon in the same place with the same label and mean completely different things — because one is a record of what was observed and the other is a prediction of what is most likely. This distinction is the single most important piece of nature-data literacy for anyone doing desk-based ecological assessment, and it underlies almost every dataset comparison on WildKnowledge.
- Surveyed data (e.g. PHI): a person, or a documented source, determined the habitat. Evidence-based.
- ML-derived data (e.g. Living England, UKCEH LCM): an algorithm inferred the most probable class from satellite imagery. Model-based.
Why it matters for nature strategy
The two carry different kinds of uncertainty, and confusing them produces predictable, expensive errors:
- A survey can be out of date; a model can be confidently wrong. Stale survey data is at least known to have been true once. A misclassified segment was never true — the model simply guessed.
- Absence means different things. No PHI polygon means "not surveyed / not a priority habitat here" — it is silence, not "nothing here". A model, by contrast, labels everywhere, so it never leaves a gap — which can lull you into thinking full coverage means full knowledge.
- Confidence is explicit vs implicit. A good ML product publishes a per-feature reliability score; a survey compilation usually does not, so you must reason about its provenance yourself.
How to read each one
When you see an ML-derived class, ask:
- What is the reliability score here? If the product publishes one (Living England does), read it. A "very low" reliability class is a hypothesis, not a fact.
- Is this class spectrally confusable? Grassland sub-types, wet vs dry habitats, and mosaics are classic misclassification zones.
- Is the feature big enough to resolve? Sub-pixel and linear features are routinely missed or merged at ~10 m resolution.
When you see surveyed data, ask:
- How old is the evidence? Currency, not method, is the survey's weak point.
- What was the source? A compilation blends sources of varying rigour.
- Does absence mean absent, or just unmapped? Usually the latter.
The practical synthesis
Neither method is "better" in the abstract — they answer different questions. Survey answers "what is confirmed here?"; ML answers "what is probably here, everywhere?" The competent desk assessment uses ML for coverage and context and survey for confidence, and — crucially — treats a high-stakes conclusion drawn from a low-confidence model as a trigger for fieldwork, not a finding.
Related datasets
- Priority Habitat Inventory — the surveyed/compiled exemplar.
- Living England — the ML-derived exemplar, notable for publishing reliability.
- UKCEH LCM — the other major ML land-cover product, useful as a contrast.
WildStack's take
The most damaging habit in desk-based ecology is treating a machine-learning land cover class as if it were a survey result — because on screen they are indistinguishable. Both are just a coloured polygon with a label. The label hides whether a human ever confirmed it.
Our discipline is to carry the method with the data: a habitat is never just "lowland meadow", it is "lowland meadow (PHI, surveyed)" or "lowland meadow (Living England, reliability: low)". Those are different claims, and a BNG baseline that flattens them into one number is quietly overstating what it knows. The reliability field exists precisely so you don't have to pretend — use it.
Official sources
- Living England: from satellite imagery to a national scale habitat map — Natural England blog
- Priority Habitats Inventory (England) — data.gov.uk
- UKCEH Land Cover Maps
Last reviewed
5 July 2026. This is a stable methodology topic — revisit mainly if the major datasets change how they express confidence, or if a new class of habitat-mapping method (beyond survey and imagery-ML) becomes mainstream.