Satellite Imagery Compressed to 64 Dimensions to Identify Tomato Fields

anselm · x · 2026-07-16

Google DeepMind's AlphaEarth Foundations can compress a year's worth of '10-meter per pixel' satellite imagery into 64 numerical values. Using this representation, a UC Davis team successfully identified all tomato fields in California.

The original text emphasizes that such tasks traditionally rely heavily on manual selection of color bands and seasonal features, requiring re-adaptation for different years or regions. The goal of the new method is to eliminate these fragile feature engineering steps, making surface crop mapping much more stable.

Original post →

More from Research

Research channel →