Deep Learning for Satellite Image Processing, Lessons Learned
summary
Deep learning is used to turn raw satellite images into production outputs using TensorFlow, Google Cloud and Spark to train and deploy models to satellite ground segments. Concrete results include a four-fold reduction in cloud-detection error in three weeks, two months to auto-generate maps from sparse OpenStreetMap labels, and denoising that matches previous state of the art; operational practices include a reproducible training version-control scheme, visual data inspection, continuous integration, and ~40× speedups from GPU runs.
AI-written summary · about
chapters (13)
- 0:00Introduction
- 3:50Deep learning revolution
- 7:00Deep neural networks basics
- 11:23Human to automation
- 15:14Cloud segmentation demo
- 20:09Deep learning history
- 22:57DeepMind object recognition
- 26:46Satellite imaging applications
- 33:29Model version control
- 37:19GPU and cloud optimization
- 40:05Adversarial and self-supervised trends
- 43:34Deep learning limits
- 45:22Tools and infrastructure
related talks
18mDeep learning for Space imagery applications
36mDeep Learning - Segmentation, Identification and Classification
15mKeynote: Incremental Deep Learning For Satellite with KubeEdge...
33mOvercoming the Barriers to Deep Learning in Production
45mIntroduction To Deep Learning
01h 36mFrom Maps to Models: An End-to-End Journey in Geospatial Machine Learning with Python
13mTaking Machine Learning to the Next Level
31mFrom Radar Echoes To Real-Time Recognition: Deep Learning on Another Planet -An Open Source Workflow