AlphaEarth embeddings map California tomato fields at 99.19% pixel accuracy
anselm · x · 2026-10-24
A UC Davis team used Google DeepMind's AlphaEarth geospatial embeddings to map California processing tomato fields, per an arXiv paper. They built a balanced dataset of 4,742 tomato and 4,742 non-tomato fields from LandIQ 2018 polygons, extracted 64-band embedding chips, and trained a U-Net on AWS SageMaker (single A10G GPU) with composite masked BCE + soft Dice loss. Monte Carlo dropout at inference (100 repeats per chip) yielded predictive mean and variance. The model hit 99.19% pixel accuracy on an independent test set. AlphaEarth compresses a year of satellite imagery for every 10m of Earth into 64 numbers, replacing fragile hand-engineered spectral features that break across seasons and regions.
More from AGI Musings
- Gary Marcus: the most dangerous people are rich, powerful and intellectually lazy — GaryMarcus · 2026-10-25
- Yacine amplifies take: in tech money no longer matters, larping collaborators are the costliest risk — yacineMTB · 2026-10-24
- AI video creators now ship in days what once took months — enggirlfriend · 2026-10-24
- Yacine on open weights: 'If your AI is open weights, your country is the product' — yacineMTB · 2026-10-24
- Researcher blasts Hinton: anthropomorphism, hype disconnect and ideological fear-mongering — gerardsans · 2026-10-24
- Grok lays out a path to billions of humanoid robots by the 2030s — Dr_Singularity · 2026-10-24