Geospatial AI combines computer vision with geographic data for powerful earth observation applications. Data sources: Sentinel-2 (free, 10m resolution, multispectral), Planet Labs (daily 3m commercial), Maxar (50cm commercial for high precision). Key tasks: 1) Image classification: classify land use / land cover (forest, water, urban, agriculture). U-Net architecture for semantic segmentation works well. Training data from Copernicus Land Monitoring Service. 2) Change detection: compare multi-temporal imagery to detect deforestation, urban growth, disaster damage. Siamese network comparing before/after pairs or differencing spectral indices. 3) Object detection: building detection, vehicle counting, ship detection. YOLO or Detectron2 fine-tuned on aerial imagery. SpaceNet dataset for training data. 4) Flood mapping: Sentinel-1 SAR (Synthetic Aperture Radar) data works through clouds. Binary classification of flooded vs non-flooded pixels using RGB + SAR fusion. Libraries: rasterio for raster data I/O, GeoPandas for vector data, sentinelhub for Sentinel API access, torchgeo for geospatial training utilities (handles geo-referenced imagery, coordinate projections). EO-learn for time series analysis of satellite data. Business applications: urban planning (building permit compliance), agriculture (crop monitoring for insurance), logistics (port activity monitoring).