Satellite Imagery to Machine-Readable Intelligence
summary
Claims a pipeline that turns multi-band, temporal Sentinel data into one embedding per pixel using the open TERA model, stores and parallel-writes chunked arrays with Zarr, and orchestrates GPU batching and durable execution with Ray. Builds D-pixel time-series (40 timestamps × multi-band) and runs inference across 12 A100 GPUs to process 40,000 km² with 60 timestamps in ~40 minutes. Produces PCA-compressed visualizations, cosine-distance change maps and masks for deforestation, roads and coastal erosion; the system scales and resumes on failures.
AI-written summary · about
chapters (13)
- 0:00Introduction to Zupole
- 1:09Geospatial AI challenges
- 4:41Satellite imagery complexity
- 9:03SAR data format
- 11:50TERA foundational model
- 15:41Pipeline bottlenecks
- 19:18Ray Data optimization
- 22:13Scaling results
- 23:15Visualizing embeddings
- 25:20Use cases: change detection
- 26:46Coastal erosion analysis
- 27:28Lessons learned
- 29:39Q&A session
related talks
36mClassification of Satellite Imagery
14mSatellite data monitoring for analytical models
15mChallenges in Practice: Building a Usable Library for Planetary-Scale Embeddings
05mAlex Kudriashova Interview
11mTerraKit: Standardising AI-Ready Geospatial Data...
41mSnakes in the Microwaves: How Python is Powering the Golden Age of SAR
10mUbiquitous Satellite Imagery of our Planet
20mOpen-source processing pipelines for Cyclone Gabrielle