Challenges in Practice: Building a Usable Library for Planetary-Scale Embeddings
summary
Pre-computed TESSERA 128-d embeddings replace a year of 10 m optical and radar imagery, trained on 800 billion non-cloudy pixel time series and stored by year (each year requires 20,000 GPU hours + 2 million CPU hours and ~200 TB online). The geotessera Python library fetches, caches, reprojects and streams embeddings as numpy arrays or GeoTIFFs; a demo shows a pixel-wise logistic classifier hitting 88% with 2 labels and 98% with 10 labels.
AI-written summary · about
chapters (6)
- 0:00Planetary-scale remote sensing challenges
- 3:07Tessa embedding model
- 5:33Embedding storage demands
- 6:15Geotessera Python library
- 7:19Demo: solar farm classification
- 11:51Q&A: interpretability & economics
related talks
39mSatellite Imagery to Machine-Readable Intelligence
11mTerraKit: Standardising AI-Ready Geospatial Data...
01h 36mFrom Maps to Models: An End-to-End Journey in Geospatial Machine Learning with Python
31mFrom Radar Echoes To Real-Time Recognition: Deep Learning on Another Planet -An Open Source Workflow
34mGeoscience Big Data Computing with Python
33mWhat and where? - Machine learning for geospatial image analysis
45mStars, Planets, and Python
17mGeoAI: a feature extraction tool set using Tensor Flow and Keras...