Habitat Loss Mapping

Blog

Attributing land-cover change to its drivers for an SDG report

Aerial view of a suburban housing development spreading into open farmland under a clear sky

The biennial update report needs a single hectare figure for land-cover change. That number gets filed without much argument. The question that follows, from reviewers and from the custodial agency if you're feeding SDG indicator 15.3.1 or 15.1.1, is what caused it. A loss of forest or grassland on the map has to be split into drivers before it means anything to a policy desk: how much is cropland expansion, how much is settlement growth, how much is mining, roads, or something else entirely.

The UNCCD's land degradation neutrality good practice guidance and the IPCC's land-use reporting framework both expect this split, usually against a short, standard list of proximate drivers. Getting from a transition matrix (forest to cropland, grassland to built-up, and so on) to a driver code per polygon is the step most national teams find genuinely hard, because the same raw transition can hide two very different causes.

Proximate drivers versus underlying drivers

It helps to keep the two apart from the start. Underlying drivers are the forces behind the scenes: population growth in a district, commodity prices, a new land tenure policy, migration out of a flooded coastal zone. You can discuss these in the narrative section of the report, but you can't code them pixel by pixel. Proximate drivers are the visible, on-the-ground actions that convert the land: clearing for crops, construction of housing or roads, extraction, forestry operations. The standard SDG and LDN driver lists work at this proximate level, usually five or six categories plus "other" or "unknown."

Your job as the analyst is to assign one proximate driver to each change polygon, or each cell in the transition matrix, consistently across the whole territory and, ideally, across the whole time series. That consistency, not the precision of any one label, is what survives peer review.

Telling agriculture and urban expansion apart in the same transition

This is where most of the disputed cases live. A parcel of grassland or secondary forest converts to bare or built surface, and on the imagery alone that can look like either new cropland being cleared or the leading edge of a town. The pixel signature right after clearing is similar in both cases; what differs is what shows up next.

A working rule set usually leans on ancillary layers rather than the spectral signal: a gazetted urban or peri-urban boundary, building-footprint or settlement layers, a buffer around existing road and utility corridors, cadastral parcel boundaries where they exist. If the converted parcel falls inside a declared urban extent or within a set distance of continuous built-up, code it as urban expansion. If it sits in a rural cadastral block with no nearby settlement growth and the post-conversion land cover tracks a cropping calendar across subsequent years, code it as agricultural expansion. Mining and quarrying concessions, where licensed, should be pulled in as a separate flag rather than left to default into one of the other two buckets, since they otherwise inflate whichever category catches the leftovers.

Where none of the ancillary evidence resolves the case, it's more defensible to carry the polygon as "driver unresolved" for that cycle than to force a guess into the agriculture or urban column. A clean unresolved share holds up under questioning. A padded agriculture or urban figure, once someone asks how it was derived, usually doesn't.

Keeping the driver breakdown stable across reporting cycles

The base layer underneath a national driver table is usually what breaks it between reporting cycles. If each cycle's land-cover change figures come from a freshly commissioned classification, with a different consultant, different class definitions, and a different comparison baseline, the driver shares move around for reasons that have nothing to do with what happened on the ground. A jump from 40% to 55% agricultural expansion between two reports is hard to explain to a reviewer if you can't first rule out a methodology change.

Anchoring the driver attribution step to a change layer built once and refreshed on the same rules every year removes that variable. The ancillary-data overlay and driver coding still sit with your team, since that judgment call depends on national context the imagery alone won't give you, but the transition matrix feeding it stops shifting underneath you.

If your driver table needs a base layer that doesn't reset every reporting cycle, that's the problem Habitat Loss Mapping is built to solve.

Get started

← Back to the blog