Why Land-Use AI Models Trained Abroad Struggle in Nigeria, and What Works Instead
Land-cover and building-detection models trained mostly on data from other regions often perform worse in Nigeria, because the landscape, the seasons, the haze and the labels differ from what the models learned. Local labels, seasonal time series and field validation are the practical fixes.
Why it happens
Training data
A review of Earth-observation datasets found that nearly 40% cover only Europe or North America, while Africa accounts for about 5%. Benchmarks of geospatial foundation models report weaker results in under-represented regions such as Africa, and models trained on one continent can perform very differently on another. A model that scores well on a benchmark has not necessarily seen anything like your study area.
The built environment
Nigerian compounds mix materials: mud-brick, zinc-roofed and unfinished concrete structures can share one compound. Google’s Open Buildings team notes that buildings the same colour as their surroundings and very small buildings are hard to detect at satellite resolution, and that large or complex buildings are sometimes split into several detections. [Optional, only if it comes from your own testing: In my own tests of an open building-detection model on Kano imagery, it missed roughly 40% of structures despite strong benchmark scores.]
Seasons, haze and cloud
Land-cover signals depend on climate and season. A smallholder mosaic of maize, sorghum, fallow and grazing can interleave at scales finer than a satellite pixel, which a model trained on large, uniform fields will misread. During the harmattan, dust haze degrades optical imagery over the north for weeks. In the south, persistent cloud causes a similar problem.
The labels
Definitions differ between products. What counts as “forest” or “cropland” in a global dataset may not match savanna woodland, farm fallow or plantation mosaics on the ground, so a model can follow its training faithfully and still be wrong locally.
What works, in order of effort
Check what already exists. Google’s Open Buildings is Africa-focused, trained on 1.75 million manually labelled buildings from imagery across the continent, and each building carries a confidence score you can filter on. GRID3 Nigeria Settlement Extents (v4.1, released August 2026, CC BY-SA 4.0) maps settlements and blocks countrywide, and its v4.0 update removed 2,269 settlements that fieldwork showed were false detections. Use these as baselines and check their licences, but treat none of them as truth.
Fine-tune on local labels. Needs vary by task, but the numbers can be modest: in a multi-continent crop-mapping study, about 100 labelled images gave high overall accuracy, and roughly 900 were needed to handle class imbalance.
Use time, not snapshots. A full year of NDVI typically separates fallow from cropped land better than any single image. Where haze or cloud blocks optical data, radar such as Sentinel-1 can fill the gaps.
Validate with local ground truth, and report accuracy by region and season instead of a single headline score. A community member who knows which plots lay fallow last season is worth more than a benchmark percentage.
The bigger point
Satellites pass over Nigeria constantly, but coverage is not capability. Capability comes from validated local labels, local expertise and models checked against Nigerian landscapes. Sentinel and Landsat imagery and tools like Python and QGIS are free, but the very-high-resolution imagery that building detection needs is not. In my view, validated Nigerian training datasets for crops, land cover and compound-level buildings are among the most valuable things a practitioner here could build.
By A. B Wayforward