## TL;DR
AI sees the planet from space -- counting ships at sea, detecting new construction, monitoring crop health, and mapping humanitarian crises. From NASA's open-source Prithvi model to commercial geospatial intelligence platforms, foundation models are democratizing satellite AI.

## Core Explanation
Satellite AI: (1) Object detection -- identify objects (ships, aircraft, vehicles, buildings). Challenge: objects are tiny (ships are 10-20 pixels at 10m resolution). Rotated bounding boxes for oriented objects; (2) Change detection -- compare images at T1 and T2, detect what changed (new construction, deforestation, disaster damage). Siamese networks, change transformers; (3) Segmentation -- classify each pixel (land cover, crop type, building footprint); (4) Foundation models -- pretrained on global satellite archives (Sentinel, Landsat, Planet). Prithvi (NASA+IBM): ViT with masked autoencoding on HLS data. Fine-tuned with 100-500 labeled examples. Clay Foundation Model (2024): open-source, community-developed geospatial foundation model.

## Detailed Analysis
Ship detection: xView and DOTA datasets -- annotated satellite images with 60+ object classes. YOLO/DETR adapted for tiny object detection via multi-scale feature pyramids. Maxar (30cm) enables vehicle-level detection. Orbital Insight: AI tracks oil storage tank levels (floating roof height from SAR), retail parking lot fullness, and construction activity from satellite time series. Descartes Labs: AI crop yield forecasting from satellite vegetation indices. MapWithAI (Meta): AI-extracted roads added to OpenStreetMap. Key challenge: cloud cover -- 67% of Earth is cloud-covered at any time. SAR (Sentinel-1) penetrates clouds, but SAR+optical fusion AI is still developing. Temporal resolution: revisit time (1-16 days) limits real-time monitoring.