A national AI map learns the land categories that planners use

A patch of trees inside a developed site creates a revealing mapping question. Should it count as vegetation, or as part of the developed parcel? Both answers can serve a purpose. An ecological inventory and a land-use account are asking different questions of the same place.

The Planning-Data-enabled Land Cover of China dataset, or PDLC, explicitly learns the planning answer. Its Scientific Data paper was published on 19 October; the map primarily represents conditions around 2021. Built-up labels can include internal vegetation, courtyards and shadows when they belong to a developed planning unit.[1]

For comparison, ESA WorldCover defines built-up cover around buildings, roads and other artificial structures, and explicitly excludes urban green areas such as parks. The same familiar category name can therefore carry different inclusion rules.[2]

Imagine a hypothetical 12-hectare developed parcel containing eight hectares of roofs and paving and four hectares of internal greenery. Reporting 12 hectares of developed land and eight hectares of hard surfaces would not be a contradiction. They are different measurements. Training an AI requires deciding which one its answers should reproduce.

Planning maps become teaching material

Using planning maps and manual checks, the team built over 330,000 labelled tiles from 15 cities. A ConvNeXt image model sees surrounding context while predicting a central region, helping it follow roads and parcel edges.[1]

PDLC supplies six broad classes: built-up land, cropland, roads, green spaces, water and others. The dataset record describes a two-metre product distributed as georeferenced TIFF files, organised by administrative areas. Version 1.1, released on 25 June, added imagery-access scripts and documentation.[3]

For a regional analyst, that combination suggests a focused use: assemble a common view of settlements, roads and farmland, then relate it to the administrative area being studied. A named category makes each coloured region usable in an account. A consistent definition makes accounts from neighbouring areas easier to compare.

The internal test achieved 88.6% accuracy on held-out tiles from the sampled cities. An independent 7,155-point assessment achieved 73%, with roads merged into built-up for that comparison.[1]

Green space combines ecological types; water’s internal-test precision was 39.2%. Those classes need specialised data for detailed ecological or hydrological measurement.[1]

A baseline with a defined meaning

The release’s administrative organisation provides an entry point for analysts choosing an area of interest. Its files encode the classes numerically; the repository lists the key. The record also directs users to external downloads for part of the national coverage. Source imagery and planning maps are not bundled with the derived classification.[3]

A useful analysis would preserve that class key beside its results. If another map calls a site vegetated while PDLC includes it within built-up land, the first question is whether the products intended to measure the same thing. Treating every disagreement as a model error would discard the purpose of the planning labels.

The broader contribution is a way to connect image recognition with the categories an institution actually needs to count. Sharper imagery helps locate boundaries; suitable teaching labels tell the model what those boundaries mean. PDLC puts both decisions into the construction of a national dataset.

Report an error or suggest a correction