Learn the model.

Exactly what the model does

This is not a predictive model in the machine-learning sense. It is a rule-based screening tool.

It divides England into 1 km hexagons, checks each hexagon against several national datasets, turns those inputs into component scores, and combines those scores into three final scenario views.

So the result is not "rewild here". It is closer to: "based on these datasets and assumptions, these places look more worth checking than most others".

Which datasets it uses

The published run uses six main inputs:

1. England boundary data: defines the study area so only England is scored.
2. CORINE land cover: used for habitat context and to estimate where existing semi-natural habitat already is.
3. Agricultural Land Classification: used as a rough proxy for lower agricultural conflict.
4. Dedicated flood data: used to capture floodplain or wetland restoration context.
5. Dedicated peat data: used to capture peat-related restoration context.
6. Bird and mammal observation records: used as a cautious biodiversity signal, based on NBN Atlas records from England.

How it turns those datasets into scores

1. It builds a national 1 km hex grid across England.
2. For each hexagon, it measures how much habitat is already there and how close the hexagon is to existing habitat.
3. It checks the dominant agricultural land grade in that hexagon.
4. It measures flood and peat presence and also how near the hexagon is to flood- or peat-related areas.
5. It counts bird and mammal records in the hexagon, then dampens the score where records are too sparse to be very trustworthy.
6. It converts those pieces into comparable component scores.
7. It combines those component scores into three final scenario scores.

What the model is actually doing under the hood

The model is not trained on past rewilding projects. It uses a fixed set of rules to compare every 1 km cell in England in the same way.

For each cell, it creates a small set of component scores. Each one captures a different part of the picture.

• Habitat share: how much of the cell is already covered by semi-natural habitat.
• Habitat distance: how near the cell is to existing habitat networks.
• Restoration opportunity: higher where a cell is near habitat but still has room to recover, rather than already being mostly habitat.
• Agricultural opportunity: higher where the dominant agricultural land grade suggests lower conflict with productive farmland.
• Flood opportunity: higher where flood-related land is present or nearby, based on the dedicated flood layer used in the canonical run.
• Peat opportunity: higher where peat-related land is present or nearby, based on the dedicated peat layer used in the canonical run.
• Biodiversity observation: higher where bird and mammal records suggest stronger biodiversity interest, but reduced where record coverage is too sparse to trust strongly.

Those components are then combined into final scenario scores. So the model is never saying "this is the correct site". It is saying "under this set of datasets and weights, this cell looks stronger or weaker than others".

What each score is really trying to capture

Habitat and restoration opportunity: the model gives more credit to places that are near existing habitat but still have room to recover. It is trying to find places that could extend or connect habitat, not just places that are already in good condition.

Biodiversity observations: the model uses bird and mammal records as a rough signal of ecological interest, but it treats low recording effort as low confidence, not proof that wildlife is absent.

Agriculture: the model treats poorer agricultural land as a rough sign that rewilding might involve less conflict with farming. This is only a proxy, not a farm business assessment.

Flood opportunity: the model combines how much flood-related area is in a cell with how near the cell is to flood-related land.

Peat opportunity: the model does something similar for peat, combining peat extent or weight with proximity.

How the published run builds each component

Habitat context: CORINE land-cover data is used to estimate where semi-natural habitat already exists. The model measures both habitat share within each cell and distance from the cell to the nearest habitat polygon.

Restoration opportunity: the model combines those habitat measures so that cells near habitat can score well, but cells already dominated by habitat do not automatically rise to the top. This is meant to favour extension, buffering, and connection, not just intact habitat alone.

Agricultural opportunity: the model takes the dominant Agricultural Land Classification grade in each cell and converts it into a 0 to 100 opportunity score. Poorer agricultural land gets a higher score because it may involve less conflict with productive farming.

Flood opportunity: the canonical run uses a dedicated flood dataset, not just land-cover classes. The model combines weighted flood-related area inside a cell with proximity to flood-related land, giving slightly more weight to area than to proximity.

Peat opportunity: the canonical run also uses a dedicated peat dataset. It combines weighted peat-related area with proximity to peat-related land, again giving more weight to area than to distance.

Biodiversity observation: bird and mammal observations are downloaded from NBN Atlas for England, filtered to recent years, then counted per cell as species richness and record count. The model scales richness, then dampens it where there are too few records to be very confident.

Boundary penalty: cells badly clipped by the England boundary or coastline are scaled down so tiny fragments do not appear artificially important.

How it reaches its final shortlist

After all the component scores are built, the model combines them in three different ways.

1. Nature-first: gives more weight to ecological upside.
2. Balanced: keeps restoration, biodiversity, flood, peat, and conflict in a more even mix.
3. Lower-conflict: gives more weight to avoiding stronger agricultural trade-offs.

The same cell can rise or fall depending on which scenario you choose. That is deliberate. The model is meant to show how priorities change the shortlist, not pretend there is one single objective answer.

From there, it ranks cells nationally, identifies strong clusters of neighbouring cells, and turns those into the shortlists, maps, and case-study outputs shown on this site.

The three scenario lenses

Each scenario uses the same component scores but changes how much weight each one gets.

• Nature-first: gives most weight to restoration opportunity and biodiversity, with meaningful additional weight for flood and peat.
• Balanced: spreads weight more evenly across restoration, biodiversity, flood, peat, and agricultural trade-off.
• Lower-conflict: gives much more weight to agricultural opportunity, so cells rise when they appear less likely to conflict with productive farmland.

In the current published run, the weights are:

• Nature-first: restoration 35%, flood 15%, peat 15%, agriculture 10%, habitat mosaic 5%, biodiversity 20%.
• Balanced: restoration 30%, flood 20%, peat 15%, agriculture 15%, habitat mosaic 5%, biodiversity 15%.
• Lower-conflict: restoration 18%, flood 15%, peat 10%, agriculture 35%, habitat mosaic 7%, biodiversity 15%.

The value judgement is built in and visible. A place can move up or down the ranking because the user's priorities have changed, not because the cell itself has changed.

What a score means

A high score is not a recommendation. It means a cell looks promising under the chosen scenario, based on the available national datasets. The next step is local review: ecology, ownership, farming context, community views, access, cost, policy fit, and delivery.

It does not prove ecological outcome, land availability, deliverability, local consent, or cost.

How to read the map

1. Start with the balanced scenario to see the current all-round shortlist.
2. Switch to nature-first to see cells that rise when ecological opportunity is weighted more strongly.
3. Switch to lower-conflict to see cells that rise when agricultural trade-off matters more.
4. Open a candidate cell and check which component scores are doing the work.
5. Treat the result as a question to investigate, not an answer to implement.