Geospatial AI: How Machine Learning Is Transforming Earth Observation and GIS

When I studied Geomatics Engineering at the University of Lagos, much of GIS work meant digitising features by hand, classifying imagery pixel by pixel with statistical methods, and spending long hours reconciling survey data with outdated base maps. The core discipline, understanding where things are and how they relate in space, has not changed. What has changed is scale.

Public missions such as the European Space Agency's Sentinel-2 and the USGS and NASA Landsat programme provide free, frequent multispectral imagery of the entire land surface, and commercial constellations add daily, high-resolution coverage. No team of analysts can look at all of it. Geospatial AI is how we turn continuous Earth observation into continuous understanding.

Semantic segmentation

Most geospatial AI tasks are segmentation problems: assign a class to every pixel, such as water, building, road, cropland or forest. The U-Net architecture, originally proposed for biomedical images, became a workhorse in remote sensing because its encoder-decoder design with skip connections preserves fine spatial detail. Variants with modern backbones remain strong baselines for building footprint extraction, road mapping and flood delineation.

Object detection and change detection

Detection models locate discrete objects, vehicles, ships, solar panels, informal structures, while change detection models compare images of the same place over time to flag new construction, deforestation or damage after a disaster. Siamese networks, which process two images with shared weights and compare their features, are a common design.

Time series models

Because satellites revisit the same location every few days, each pixel becomes a time series. Temporal convolutional networks, recurrent networks and transformers learn seasonal signatures, which is how crop types are distinguished even when they look similar in a single image.

Geospatial foundation models

The biggest recent shift is foundation models pre-trained with self-supervised learning on massive amounts of unlabelled imagery. The IBM and NASA Prithvi model, trained on Harmonized Landsat Sentinel-2 data, can be fine-tuned for flood mapping, burn scar detection and crop classification with relatively little labelled data. General-purpose vision models such as Meta's Segment Anything are also being adapted to accelerate annotation of aerial and satellite imagery.

• Near real-time land cover. Google and the World Resources Institute's Dynamic World dataset, described in Scientific Data, provides 10-metre land cover predictions from Sentinel-2 imagery as new scenes arrive, rather than a static map updated every few years.
• Urban planning and informal settlements. In fast-growing cities like Lagos, official maps lag reality. AI-extracted building footprints and road networks help planners, utilities and humanitarian organisations understand where people actually live.
• Disaster response. Synthetic aperture radar can see through clouds, and ML models rapidly map flood extents, which is essential during the rainy season floods that regularly affect parts of Nigeria.
• Environmental monitoring. Deforestation alerts, mangrove loss, gas flaring detection and coastal erosion tracking all rely on automated analysis of imagery time series.
• Infrastructure and addressing. Automated mapping supports last-mile logistics, electrification planning and property registration in places where formal addresses are incomplete.

Geospatial AI is powerful, and that power cuts both ways.

• Geographic bias. Many training datasets over-represent North America, Europe and China. A building detector trained on suburban American houses may struggle with dense, irregular informal settlements or with compound houses common in West Africa.
• Label quality. OpenStreetMap and other volunteered data are invaluable but uneven. Models learn the errors in their labels.
• Spatial autocorrelation in evaluation. Randomly splitting pixels into train and test sets leaks information because nearby pixels are similar. Use spatially blocked cross-validation, or you will overestimate accuracy, a mistake I still see in published work.
• Privacy and surveillance. High-resolution imagery combined with AI can reveal sensitive information about individuals, communities and critical infrastructure. Teams should consider how outputs could be misused, for example in forced evictions or targeting.
• Accuracy communication. A coloured map looks authoritative. Always publish accuracy assessments, confusion matrices and known failure modes alongside the map.

For engineering teams, especially in emerging markets where budgets are tight, a pragmatic pipeline looks like this:

1. Define the decision. What will someone do differently with this map? That determines resolution, update frequency and acceptable error.
2. Source data cleverly. Start with free Sentinel and Landsat imagery via cloud platforms that host analysis-ready data. Buy commercial imagery only for tasks that genuinely need sub-metre detail.
3. Label locally. Combine field surveys, local knowledge and assisted annotation tools. A small, high-quality, representative local label set is worth more than a large foreign one.
4. Fine-tune a foundation model. Rather than training from scratch, fine-tune a pre-trained geospatial model on your labels.
5. Evaluate spatially. Hold out whole regions or tiles for testing.
6. Serve results as standard spatial data. Output cloud-optimised GeoTIFFs, vector tiles or GeoJSON, store vectors in PostGIS and expose them through OGC-compliant APIs so that GIS users can consume them in QGIS or web maps.
7. Automate updates. Schedule inference as new imagery arrives and track changes over time.

Python's geospatial ecosystem, rasterio, GeoPandas, xarray and libraries such as TorchGeo, makes this accessible to backend engineers who already know Django or FastAPI. A Django backend with GeoDjango and PostGIS is, in my experience, an excellent foundation for serving AI-derived spatial products.

We are moving towards a world where asking a question of the planet, where did new buildings appear in this ward last quarter? or which farms in this district show water stress?, can be answered in near real time through a combination of foundation models and natural language interfaces. Large language models that can call geospatial tools and query spatial databases are already being prototyped.

For Africa, this is a chance to leapfrog decades of incomplete mapping. But it requires local geospatial talent, local data and local ownership. As someone who mentors developers through BuildWithMMO, I would love to see many more African engineers combine GIS fundamentals with modern ML; it is one of the most valuable and underserved skill combinations on the continent.

• Free, frequent satellite imagery plus deep learning has made continuous mapping of the Earth practical.
• Segmentation (U-Net and successors), change detection, time series models and geospatial foundation models are the core techniques.
• Geographic bias, label quality and spatial leakage in evaluation are the most common technical pitfalls.
• Consider privacy and misuse carefully, and always publish accuracy assessments with your maps.
• Fine-tune foundation models on local labels and serve outputs through standard GIS formats and PostGIS-backed APIs.

1. Sentinel-2 — European Space Agency
2. Landsat Missions — U.S. Geological Survey
3. U-Net: Convolutional Networks for Biomedical Image Segmentation — arXiv
4. Foundation Models for Generalist Geospatial Artificial Intelligence — arXiv
5. Segment Anything — arXiv
6. Dynamic World, Near real-time global 10 m land use land cover mapping — Scientific Data

Cover image: original illustration created for this article. Cited works remain the property of their respective authors and publishers.