Democratizing Climate Science
summary
SpaceML builds a production reverse‑image search over NASA GIBS covering 20+ years and petabytes of imagery, using a GIBS downloader to tile multi‑band data, GPU‑accelerated feature extraction with NVIDIA DALI, 128‑dimension 128‑byte embeddings, and approximate nearest neighbor indexes (FAISS, SCANN, Annoy). Cost estimates: $8/day (250m), $76/day (30m), $100k to reprocess 20 years. A search API, swipe labeler and Chrome plugin are live and the system is running in production.
AI-written summary · about
chapters (18)
- 0:00Intro
- 0:12Wildfires
- 1:10Data
- 3:56Unlabeled Data
- 5:01Unbalanced Data
- 6:32NASA
- 7:08Technology Readiness Level
- 8:35What is SpaceML
- 9:13How does SpaceML work
- 9:52What is an embedding
- 12:10Approximate nearest neighbor search
- 13:40Data size
- 17:10Cost
- 18:20Design Goals
- 20:42Architecture
- 30:14Search API
- 35:45Labels
- 37:47Team
related talks
14mSatellite data monitoring for analytical models
31mFrom Radar Echoes To Real-Time Recognition: Deep Learning on Another Planet -An Open Source Workflow
28mAdvancing the future of space in the cloud
18mDeep learning for Space imagery applications
34mGeoscience Big Data Computing with Python
09mGen AI Workflows for Climate Risk Modeling in Food Systems
45mFrom Radar Echoes To Real-Time Recognition: Deep Learning on Another Planet an Open Source Workflow
28mKubernetes Meets Climate Science: Building Large-scale Feature Detection From Climate Data Records