How AI is closing the 3% cadastre gap in developing nations

A land title system is only as complete as the parcels it knows about. In mature registries the coverage gap is negligible, but across much of sub-Saharan Africa and South-East Asia an estimated three percent of occupied land never made it into the digital cadastral database. These parcels are farmed, built on, bought, and inherited — they simply do not exist on paper, which means their occupants cannot borrow against them, defend them in a dispute, or be taxed fairly on them.

Closing that gap the traditional way means sending surveyors into the field to walk boundaries one at a time. It is accurate and it is unaffordable at national scale. The question we set out to answer was whether earth observation plus machine learning could narrow the search space enough that the expensive human step is spent only where it counts.

Why the gap is invisible to conventional tools

A missing parcel does not announce itself. There is no record to flag as incomplete, no attribute that reads "unregistered". The only signal is negative space: a patch of clearly occupied land — a compound, a cultivated field, a subdivided block — that has no corresponding polygon in the DCDB.

Detecting negative space is exactly the kind of task humans do well and rule-based GIS does badly. A hard-coded overlay of imagery against the cadastre generates thousands of false positives from roads, water bodies, communal land, and imagery artefacts. Without a way to rank those candidates, an analyst queue drowns before a single real gap is confirmed.

Contrastive learning against the existing cadastre

Rather than train a detector to recognise "a parcel" in the abstract, we train it to recognise parcels that look like the ones already in your registry. A contrastive model learns an embedding in which a satellite-detected boundary sits close to genuine DCDB parcels of the same tenure character and far from roads, rivers, and noise.

The existing cadastre becomes the label set. In a district where registered parcels are typically quarter-hectare residential blocks with orthogonal boundaries, the model learns that shape and scale; in a district of long agricultural strips it learns those. This is what keeps the false-positive rate low enough that the output is a review list, not a haystack.

• Segment candidate boundaries from high-resolution imagery.
• Embed each candidate and every registered parcel into the same learned space.
• Rank candidates by distance to the nearest genuine parcel cluster.
• Emit only high-confidence, geometrically plausible gaps for review.

Keeping a surveyor in the loop

Nothing the pipeline produces is treated as a registered parcel. Every candidate is a proposal, tagged with its confidence, the imagery date it was derived from, and the DCDB parcels it neighbours. A cadastral officer confirms, adjusts, or rejects it, and only a confirmed geometry is written back through the Cadastre Engine with full provenance.

This matters legally as much as technically. A boundary that will underpin a title cannot originate from a model with no human accountable for it. The AI narrows a national problem to a reviewable shortlist; the surveyor still owns the decision.

What the economics look like

In pilot districts the pipeline reduced the area a field team had to physically visit by an order of magnitude, because the model retired the obvious non-parcels and the confidently-registered land before anyone left the office. The saving is not that surveyors are replaced — it is that their time is spent only on the ambiguous three percent instead of re-walking the ninety-seven that were already correct.