Brainfood: Mapping edition

  1. What’s an embedding? A single spot on Earth might have dozens of pixels across multiple spectral bands (visible light, infrared, radar), observed repeatedly over time. That’s a lot of raw data, and most machine learning models don’t work well with raw pixels directly: too much redundancy and noise. An embedding model (like Google DeepMind’s Alpha Earth Foundation) is trained to compress all of that into a fixed-length vector that still preserves the underlying signal. Things like “this looks like dense broadleaf canopy” or “this has the seasonal greening pattern of a palm plantation” get encoded in that vector space, even though the model never explicitly labels them that way. You can then just run a relatively simple classifier on top of the embeddings, because the hard work of “understanding” the imagery was already done upstream. Neat.
  2. No word on embeddings, but the Land & Carbon Lab writeup of the dataset walks through why it took so long to build: livestock census data is collected at wildly different administrative scales and intervals, and the workflow that turns that mess into consistent annual maps is as much the story as the maps themselves.

Leave a Reply

Your email address will not be published. Required fields are marked *