GeoAI is the application of artificial intelligence — chiefly machine learning and deep learning — to geospatial data and problems. It spans extracting features from imagery, classifying land cover, predicting where things will happen (crime, disease, deforestation, demand), and increasingly, reasoning over spatial data with foundation models. The geography adds structure that generic ML ignores at its peril: spatial autocorrelation, scale, and the tyranny of geography-specific training data.
The GeoAI problem families
Recognition tasks dominate: detecting and delineating buildings, roads, trees, ships and change from imagery — computer vision applied to the Earth. Prediction tasks come second: risk and propensity surfaces learned from historical patterns and spatial covariates. A third family accelerates simulation and interpolation — ML emulating physical models at a fraction of their cost.
Each family inherits classic ML discipline — training data, validation, drift — plus spatial complications: models trained in one geography routinely fail in another, and spatially naive validation flatters performance.
Why spatial data breaks naive ML
Random train–test splits leak information when nearby, near-identical samples land on both sides — spatial cross-validation with geographic blocks is the honest alternative. Class imbalance is endemic (change is rare; most land is background). And covariate shift is geographic reality: roofs in Jakarta do not look like roofs in Zurich.
The craft is therefore as much data engineering as modelling: representative sampling across geographies, hard-negative mining, and evaluation that mimics deployment.
From model to product
A model is not a deliverable. Production GeoAI wraps it in pipelines — imagery ingestion, tiling, inference at scale, vectorisation, and human-in-the-loop review that catches the failures automation cannot see. Confidence scores accompany outputs; feedback loops retrain on corrected errors.
The mature pattern is AI-plus-QA: machines for scale, expert reviewers for judgement, with the division of labour tuned to the accuracy the contract demands.