Methodology
Probabilistic GeoML
Bring spatial structure — proximity and autocorrelation — into Bayesian deep‑learning models built for land‑use prediction.
01
The project
Denmark is acutely exposed to a changing climate. A long coastline, low‑lying terrain, and intensively managed land leave it vulnerable to flooding, drought, and coastal degradation — yet precise, uncertainty‑aware forecasts of how its land use will respond are still missing. Anticipating that change is essential for proactive, sustainable planning.
DK‑Future develops a new class of probabilistic geographical machine‑learning models — probabilistic GeoML — that integrate spatial structure with Bayesian deep learning. Combining historical earth observation, climate projections, socio‑economic signals, and legislative frameworks, the models forecast Danish land use at horizons of 2030, 2050, and 2100. Crucially, they don't produce a single forecast: they output probability distributions over alternative land‑use scenarios, surfacing the low‑probability but high‑impact shifts that decision‑makers need to plan for.
The work is committed to open‑science standards, so that authorities, researchers, and the public can build on its data and models.
Programme
Villum Synergy
Funding
DKK 4.48M
Duration
2026–2028
Host
Aalborg University
Supported by Villum Fonden
DK‑Future is funded through the Villum Synergy programme,
which backs interdisciplinary data‑science research — here, a synergy of
probabilistic machine learning and geographic science.