Geospatial AI applies machine learning to location data: satellite and aerial imagery, maps, GPS traces and boundary data.
Data Sources
- Satellite imagery with varying resolution, frequency and bands (visible, infrared, radar).
- Aerial and drone imagery for high detail.
- Vector data: boundaries, roads and buildings, such as the Natural Earth data on glitchdata's hub.
- GPS and mobility traces.
Common Tasks
- Land-use and land-cover classification.
- Detecting buildings, roads and changes over time.
- Monitoring crops, forests, water and disasters.
- Estimating population or economic activity.
- Optimising routes and site selection.
Techniques
Convolutional networks and vision transformers for imagery, often pretrained on satellite data; spatial statistics and gradient boosting for tabular location features.
Pitfalls
- Spatial autocorrelation: nearby areas are similar, so random train/test splits leak information. Split by region.
- Coordinate systems: mixing projections produces subtle errors.
- Clouds and seasons change what imagery shows.
- Resolution limits determine what can be detected.
- Privacy: location traces can identify individuals.
Tools
GeoPandas and QGIS for vector data; rasterio and GDAL for imagery; cloud platforms for large-scale imagery analysis.
Validate on the Ground
Where possible, check predictions against ground truth in several regions, not just the one used for training.